TF_Output inputs[2];
    TF_Output outputs[1];
    TF_Output grad_outputs[2];
    TF_Output expected_grad_outputs[2];

    BuildSuccessGraph(inputs, outputs);
    BuildExpectedGraph(grad_inputs_provided, expected_grad_outputs);

    AddGradients(grad_inputs_provided, nullptr, inputs, 2, outputs, 1,
                 grad_outputs);

  }

  std::vector<AbstractTensorHandle*> outputs(1);
  status_ = RunModel(ReluGradModel, immediate_execution_ctx_.get(), {Y.get()},
                     absl::MakeSpan(outputs), UseFunction());
  ASSERT_EQ(errors::OK, status_.code()) << status_.message();
  ASSERT_NO_FATAL_FAILURE(CheckTensorValue(outputs[0], {0.0f}, /*dims*/ {},
                                           /*abs_error*/ 0));
  outputs[0]->Unref();
}

// being equivalent to the following python code.
//
// if use_function:
//   outputs = tf.function(model)(inputs)
// else:
//   outputs = model(inputs)
Status RunModel(Model model, AbstractContext* ctx,
                absl::Span<AbstractTensorHandle* const> inputs,
                absl::Span<AbstractTensorHandle*> outputs, bool use_function);

Status BuildImmediateExecutionContext(bool use_tfrt, AbstractContext** ctx);

 public:
  explicit ReluGradientFunction(vector<AbstractTensorHandle*> f_outputs)
      : forward_outputs_(f_outputs) {
    for (auto output : forward_outputs_) {
      if (output) {
        output->Ref();
      }
    }
  }

  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {

  }
  return outputs;
}

void TFOutputsFromOutputs(const std::vector<tensorflow::Output>& outputs,
                          TF_Output* tf_outputs) {
  for (int i = 0; i < outputs.size(); i++) {
    tf_outputs[i].oper = ToOperation(outputs[i].node());
    tf_outputs[i].index = outputs[i].index();
  }
}

                                    vector<AbstractTensorHandle*> f_outputs)
      : forward_inputs_(f_inputs), forward_outputs_(f_outputs) {
    for (auto input : forward_inputs_) {
      if (input) {
        input->Ref();
      }
    }
    for (auto output : forward_outputs_) {
      if (output) {
        output->Ref();
      }
    }
  }

  Status Compute(AbstractContext* ctx,

      for (int i = 0; i < model_outputs.size(); i++) {
        if (model_outputs[i]) {
          output_list.outputs.emplace_back(model_outputs[i]);
          output_list.expected_num_outputs += 1;
        } else {
          null_indices.insert(i);
        }
      }
      TF_RETURN_IF_ERROR(dyn_cast<tracing::TracingContext>(func_ctx.get())
                             ->Finalize(&output_list, &func));

    }
    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) {

                      absl::Span<AbstractTensorHandle*> outputs) {
  std::vector<AbstractTensorHandle*> temp_outputs(2);
  TF_RETURN_IF_ERROR(
      ops::IdentityN(ctx, inputs, absl::MakeSpan(temp_outputs), "IdentityN"));
  // Although, `ops::IdentityN` returns 2 tensors, the first tensor isn't needed
  // for computing gradient so we could safely drop it.
  outputs[0] = temp_outputs[1];
  temp_outputs[0]->Unref();
  return absl::OkStatus();
}

  // device if any of their inputs are on that custom device, but custom devices
  // are free to set a bad status in order to require explicit placement.
  void (*execute)(const TFE_Op* op, int* num_outputs,
                  TFE_TensorHandle** outputs, TF_Status* s, void* device_info);

  // Method to delete a device.
  void (*delete_device)(void* device_info);

  // Implements TFE_CreatePackedTensorHandle when one of `handles` is on this