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,

        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;
    }

  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 =

  void MarkAsResult(AbstractTensorHandle* gradient) const override;

  void DeleteGradient(AbstractTensorHandle* gradient) const override;

 private:
  // The context where the aggregation op `Add` is to be created.
  AbstractContext* ctx_;
};

// Returns the number of elements in the gradient tensor.
int64_t TapeVSpace::NumElements(AbstractTensorHandle* tensor) const {

      "      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'"

  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.

  rewrite_options->mutable_scoped_allocator_opts()->add_enable_op(
      "CollectiveReduce");

  if ((tensorflow::IsGoogleCudaEnabled() || tensorflow::IsBuiltWithROCm()) &&
      num_virtual_gpus > 0) {
    tensorflow::GPUOptions* gpu_options =
        server_def.mutable_default_session_config()->mutable_gpu_options();
    auto virtual_devices =
        gpu_options->mutable_experimental()->add_virtual_devices();

  }

  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    // TODO(vnvo2409): Add control dependency
    /* Given upstream grad U and a Log1p op: Y = log(1 + X), the gradients are:
     *
     *    dX = U / (1 + X)
     *
     */

    AbstractTensorHandle* upstream_grad = grad_outputs[0];