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) {

                 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");

           {"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()))

  NodeDef node_def;
  tensorflow::ImmediateExecutionOperation* op = tensorflow::unwrap(tfe_op);
  node_def.set_name(op->Name());
  node_def.set_op(op->Name());
  for (int i = 0; i < num_inputs; ++i) {
    node_def.add_input("dummy_input");
  }
  OperationFromInterface(op)->Attrs().FillAttrValueMap(node_def.mutable_attr());

  const tensorflow::OpRegistrationData* op_reg_data;
  status->status =

    add_output1 = TF_OutputListGet(add_outputs, 0);
    TF_DeleteOutputList(add_outputs);
  }

  // Same with a second "Add" computing `arg1 + arg1`.
  TF_AbstractTensor* add_output2;
  {
    // Build an abstract operation, inputs and output.
    auto* add_op = TF_NewAbstractOp(graph_ctx);
    TF_AbstractOpSetOpType(add_op, "Add", s);
    ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);
    TF_AbstractOpSetOpName(add_op, "my_add", s);

  EXPECT_EQ(feed, add_in_0.oper);
  EXPECT_EQ(0, add_in_0.index);
  TF_Output add_in_1 = TF_OperationInput(TF_Input{add, 1});
  EXPECT_EQ(three, add_in_1.oper);
  EXPECT_EQ(0, add_in_1.index);
  EXPECT_EQ(0, TF_OperationOutputNumConsumers(TF_Output{add, 0}));
  EXPECT_EQ(0, TF_OperationNumControlInputs(add));
  EXPECT_EQ(0, TF_OperationNumControlOutputs(add));

    for (int i = 0; i < 100000; ++i) {
      TFE_Op* add_op_dummy = AddOp(ctx, m, m);
      TFE_OpSetDevice(add_op_dummy, cpu_device_name.c_str(), status);
      ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
      TFE_TensorHandle* dummy = nullptr;
      TFE_Execute(add_op_dummy, &dummy, &num_retvals, status);
      ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);