{x1.get(), x2.get()}, absl::MakeSpan(outputs), UseFunction());
  ASSERT_EQ(errors::OK, status_.code()) << status_.message();
  EXPECT_EQ(outputs[0], nullptr);
  ASSERT_NO_FATAL_FAILURE(CheckTensorValue(outputs[1], {1.0f}, /*dims*/ {},
                                           /*abs_error*/ 0));
  outputs[1]->Unref();
}

#ifdef PLATFORM_GOOGLE
INSTANTIATE_TEST_SUITE_P(

                         tensorflow::FunctionLibraryDefinition* flib_def,
                         std::vector<std::string>* control_ret_node_names,
                         bool* control_rets_updated) override {
    // Inject failure to function instantiation if finding a node that contains
    // the given node name (error_node_) and requested device (error_device_).
    for (const auto node : graph->get()->nodes()) {

    // would otherwise deadlock).
    if (TF_GetCode(status) != TF_OK &&
        (first_bad_status == nullptr
         // Prefer propagating non-cancellation related statuses to avoid
         // shadowing the original failure.
         || TF_GetCode(first_bad_status.get()) == TF_CANCELLED)) {
      first_bad_status.reset(TF_NewStatus());
      TF_SetStatus(first_bad_status.get(), TF_GetCode(status),
                   TF_Message(status));

        immediate_execution_ctx_.get(), X_vals, X_dims, 2, &X_raw);
    ASSERT_EQ(errors::OK, status_.code()) << status_.message();
    X.reset(X_raw);
  }

  ASSERT_NO_FATAL_FAILURE(CompareNumericalAndAutodiffGradients(
      ReluModel, ReluGradModel, immediate_execution_ctx_.get(), {X.get()},
      UseFunction()));

  // Mathematically, Relu isn't differentiable at `0`. So `gradient_checker`

      EXPECT_FALSE(found_scalar_const);
      found_scalar_const = true;
    } else if (IsAddN(n, 2)) {
      EXPECT_FALSE(found_add);
      found_add = true;
    } else {
      ADD_FAILURE() << "Unexpected NodeDef: " << n.DebugString();
    }
  }
  EXPECT_TRUE(found_placeholder);
  EXPECT_TRUE(found_scalar_const);
  EXPECT_TRUE(found_add);

  // Add another oper to the graph.