// TODO(b/316226111): the printer may not guarantee the same order of fields. Rewrite the checks to be less sensitive to proto serialization formats.
// -----
// Non-SPMD: one input and one output
//
// CHECK-LABEL: func.func @serving_default(%arg0: tensor<1x3xf32>) -> tensor<1x3xf32> {
// CHECK-NEXT:  "tf.IfrtCall"(%arg0)
// CHECK-SAME:       {program_id = [[PROGRAM_ID:.*]] : i64, variable_arg_indices = []}
// CHECK-SAME:       (tensor<1x3xf32>) -> tensor<1x3xf32>

          converted_model2.signatures._signatures.keys(), {'serving_default'}
      )

      # Expect two models to produce the same results.
      test_data = ops.convert_to_tensor(
          np.random.uniform(low=0, high=64, size=(32)).astype(
              dtypes.int32.as_numpy_dtype
          )
      )
      new_outputs_1 = converted_model1.signatures['serving_default'](
          input_tensor=test_data

// RUN: tf-tfrt-opt -split-input-file -tf-mlrt-rewrite-ifrt-load-variable %s | FileCheck %s

// Variable is used by both CPU and TPU
//
// CHECK-LABEL: func @serving_default(%arg0: tensor<1x3xf32>) -> tensor<1x1xf32>
// CHECK-NEXT:    [[HANDLE:%.*]] = "tf.VarHandleOp"()
// CHECK-NEXT:    [[ARRAYKEY:%.*]], [[FURTURE:%.*]] = "tf_mlrt.tf_ifrt_load_variable"([[HANDLE]])

  func.func @serving_default(%arg0: tensor<1x20x20x28xf32> {tf_saved_model.index_path = ["a"]}) -> (tensor<1x40x40x28xf32> {tf_saved_model.index_path = ["b"]}) attributes {tf.entry_function = {control_outputs = "", inputs = "c:0", outputs = "d:0"}, tf_saved_model.exported_names = ["serving_default"]} {
      %c = stablehlo.constant dense<40> : tensor<2xi32>

  func.func @serving_default(%arg0: tensor<1x32x32x128xf32> {tf_saved_model.index_path = ["a"]}) -> (tensor<1x64x64x128xf32> {tf_saved_model.index_path = ["b"]}) attributes {tf.entry_function = {control_outputs = "", inputs = "c:0", outputs = "d:0"}, tf_saved_model.exported_names = ["serving_default"]} {
      %0  = "tf.Const"() {value = dense<[56, 904]> : tensor<2xi32>} : () -> tensor<2xi32>

    // If the signature_key has not been explicitly specified, use the default
    // value of "serving_default".
    const std::string signature_key = dataset_config.has_signature_key()
                                          ? dataset_config.signature_key()
                                          : "serving_default";
    if (repr_dataset_file_map.contains(signature_key)) {
      return absl::InvalidArgumentError(

  // LINT.IfChange(savedmodel_to_stablehlo)
  m.def(
      "savedmodel_to_stablehlo",
      [](absl::string_view input_path,
         const std::vector<std::string>& exported_model_signatures =
             {"serving_default"},
         const std::vector<std::string>& tag_names = {"serve"},
         absl::string_view input_arg_shapes_str = "") -> py::bytes {
        auto module_bytecode =

  EXPECT_THAT(FindMainFuncOp(*module_op), IsNull());
}

TEST_F(FindMainFuncOpTest, ReturnsServingDefaultFuncOp) {
  constexpr absl::string_view kModuleWithServingDefaultFunc = R"mlir(
    module {
      func.func @serving_default() -> () {
        return
      }
    }
  )mlir";

  OwningOpRef<ModuleOp> module_op =
      ParseModuleOpString(kModuleWithServingDefaultFunc);
  EXPECT_THAT(*module_op, NotNull());

// CHECK: return %[[CUSTOM_AGGREGATOR_1]] : tensor<1x3xf32>
// CHECK: }
// CHECK: }

// -----

// Tests that stablehlo op serialization also works for the "serving_default"
// function.

func.func @serving_default(%arg0: tensor<1x4xf32>) -> tensor<1x3xf32> {
  %0 = stablehlo.constant dense<1.0> : tensor<4x3xf32>

  TF_SignatureDefFunction* serving_default =
      TF_GetSavedModelSignatureDefFunction(saved_model, "serving_default",
                                           status);
  EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);

  TF_SignatureDefFunctionMetadata* metadata =
      TF_SignatureDefFunctionGetMetadata(serving_default);

  const TF_SignatureDefParamList* args =