return absl::OkStatus();
  }
  ~SparseSoftmaxCrossEntropyWithLogitsGradientFunction() override {}

 private:
  vector<AbstractTensorHandle*> forward_outputs_;
};

// TODO(vnvo2409): Add python test
class BiasAddGradientFunction : public GradientFunction {
 public:
  explicit BiasAddGradientFunction(AttrBuilder f_attrs)
      : forward_attrs_(f_attrs) {}

  Status Compute(AbstractContext* ctx,

  server_def.set_protocol("grpc");
  server_def.set_job_name(job_name);
  server_def.set_task_index(0);
  tensorflow::ClusterDef* cluster_def = server_def.mutable_cluster();
  tensorflow::JobDef* job_def = cluster_def->add_job();
  job_def->set_name(job_name);
  for (int i = 0; i < num_tasks; i++) {
    int port = tensorflow::testing::PickUnusedPortOrDie();
    job_def->mutable_tasks()->insert(

        return;
      }
      // We have a convoluted scheme here: Using the C++ graph construction API
      // to add potentially many nodes to the graph without running the checks
      // (such as uniqueness of the names of nodes) we run with other functions
      // that add a node to the graph (like TF_FinishOperation).
      if (!g->name_map.insert(std::make_pair(n->name(), n)).second) {

      "      name: 'add1'"
      "      op: 'Add'"
      "      input: 'read0:value:0'"
      "      input: 'read1:value:0'"
      "      attr {"
      "        key: 'T'"
      "        value {"
      "          type: DT_FLOAT"
      "        }"
      "      }"
      "    }"
      "    node_def {"
      "      name: 'add2'"
      "      op: 'Add'"
      "      input: 'add1:z:0'"

      TF_NewFunctionLibraryDefinition(g_buf, status);

  TF_LookUpOpDef(func, "Add", op_buf, status);
  string actual_string(reinterpret_cast<const char*>(op_buf->data),
                       op_buf->length);
  ASSERT_EQ(TF_OK, TF_GetCode(status));

  const OpDef* expected_op_def;
  TF_ASSERT_OK(OpRegistry::Global()->LookUpOpDef("Add", &expected_op_def));
  string expected_serialized;

                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    for (int i = 0; i < grad_outputs.size(); i++) {
      auto grad_input = grad_outputs[i];
      // TODO(srbs): Should we add a copy contructor to AbstractTensorHandle
      // that takes care of this similar to `Tensor`?
      if (grad_input) {
        grad_input->Ref();
      }
      grad_inputs[i] = grad_input;
    }

Status AddModel(AbstractContext* ctx,
                absl::Span<AbstractTensorHandle* const> inputs,
                absl::Span<AbstractTensorHandle*> outputs) {
  return ops::AddV2(ctx, inputs[0], inputs[1], &outputs[0], "Add");
}

Status ExpModel(AbstractContext* ctx,
                absl::Span<AbstractTensorHandle* const> inputs,
                absl::Span<AbstractTensorHandle*> outputs) {
  return ops::Exp(ctx, inputs[0], &outputs[0], "Exp");

    std::string second_device =
        TFE_TensorHandleBackingDeviceName(components[1].get(), status.get());
    ASSERT_EQ(underlying_devices[1], second_device);
  }

  // Add a parallel tensor with different values on each device to the variable.
  {
    TensorHandlePtr value_one(FloatTensorHandle(3., status.get()));
    TensorHandlePtr value_two(FloatTensorHandle(-2., status.get()));

  if (num_opers == -1) {
    for (const Node* node : fn_body->graph.op_nodes()) {
      const auto& iter = input_nodes.find(node);
      if (iter == input_nodes.end()) {
        // This node is not referenced in inputs. Add it to the body.
        body_nodes->push_back(node);
      } else {
        // This node is referenced in inputs. Currently, we place an
        // artificial restriction and require that when num_opers=-1, such

           {"x"},
           {{"T", DT_FLOAT},
            {"group_size", static_cast<int>(devices.size())},
            {"group_key", 0},
            {"instance_key", 0},
            {"merge_op", "Add"},
            {"final_op", "Id"},
            {"subdiv_offsets", std::vector<int>()}},
           /*dep=*/{"assert"}},
      });
  TF_ASSERT_OK(ContextFromInterface(unwrap(context.get()))