}
    inputs_.push_back(t->output_);
    *output = tensorflow::down_cast<TracingTensorHandle*>(outputs[0]);
    return absl::OkStatus();
  }

  Status Finalize(OutputList* outputs, AbstractFunction** f) override {
    std::vector<TF_Output> graph_outputs;
    graph_outputs.reserve(outputs->outputs.size());
    for (auto* abstract_output : outputs->outputs) {

}

class NaryGradTest : public ::testing::Test {
 protected:
  NaryGradTest() : scope_(Scope::NewRootScope().WithDevice("/cpu:0")) {}

  void RunTest(const OutputList& xs, const std::vector<TensorShape>& x_shapes,
               const OutputList& ys, const std::vector<TensorShape>& y_shapes) {
    TF_ASSERT_OK(scope_.status());
    float max_error;
    TF_ASSERT_OK((ComputeGradientError<float, float, float>(

  GraphDef graph_def;
  Scope scope = Scope::NewRootScope();
  Output a = ops::Const(scope.WithOpName("a"), {10.0f, 10.0f}, {2});
  Output axis = ops::Const(scope.WithOpName("axis"), 0, {});
  OutputList split = ops::Split(scope.WithOpName("split"), axis, a, 2).output;
  Output b = ops::Const(scope.WithOpName("b"), 10.0f, {});
  Output c = ops::Mul(scope.WithOpName("c"), split[1], b);

      NewInternalScope(&parent->graph, &status->status, &parent->refiner)
          .NewSubScope(params->name);

  const int first_new_node_id = parent->graph.num_node_ids();

  tensorflow::OutputList loop_outputs;
  status->status = tensorflow::ops::BuildWhileLoop(
      scope, OutputsFromTFOutputs(parent_inputs, num_loop_vars), cond_fn,
      body_fn, params->name, &loop_outputs);