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;

    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

  summary = "{";
  std::vector<std::string> summarized_devices = device_.SummarizeDeviceNames();
  for (int component_index = 0; component_index < tensors_.size();
       ++component_index) {
    // TODO(allenl): Add a C API for summarizing tensors. Currently custom
    // devices limiting themselves to a C API (for ABI compatibility) would need
    // to implement summarization for component tensors themselves.

    ParseGCSPath(test_dir, true, &bucket, &object, status);
    if (TF_GetCode(status) != TF_OK) {
      TF_DeleteStatus(status);
      return "";
    }
    TF_DeleteStatus(status);

    // We add a random value into `test_dir` to ensures that two consecutive
    // runs are unlikely to clash.
    std::random_device rd;
    std::mt19937 gen(rd());
    std::uniform_int_distribution<> distribution;

  TFE_OpSetAttrInt(op.get(), "group_size", group_size);
  TFE_OpSetAttrInt(op.get(), "group_key", 0);
  TFE_OpSetAttrInt(op.get(), "instance_key", 0);
  const std::string merge_op("Add");
  TFE_OpSetAttrString(op.get(), "merge_op", merge_op.c_str(),
                      merge_op.length());
  const std::string final_op("Id");
  TFE_OpSetAttrString(op.get(), "final_op", final_op.c_str(),