}
}

absl::optional<std::vector<std::unique_ptr<ParallelTensor>>>
ParallelDevice::Join(
    const std::vector<PartialTensorShape>& expected_output_shapes,
    TF_Status* status) const {
  absl::optional<std::vector<std::unique_ptr<ParallelTensor>>> result;
  // Compute per-device per-output tensors
  std::vector<std::vector<TensorHandlePtr>> per_device_output_tensors;

#include "tensorflow/core/platform/status.h"

using tensorflow::string;

namespace {

std::vector<int64_t> TensorShapeAsVector(const tensorflow::TensorHandle& handle,
                                         absl::Status* status) {
  std::vector<int64_t> shape;
  int rank = -1;
  *status = handle.NumDims(&rank);
  if (!status->ok()) {
    return shape;
  }
  shape.reserve(rank);

  absl::Status Shape(const std::vector<int64_t>** shape) const;
  TF_DataType dtype() const { return dtype_; }

  // Sets its output argument to a summary of the values of this tensor on every
  // component device.
  absl::Status SummarizeValue(std::string& summary);

  std::vector<TensorHandlePtr> release_tensors() { return std::move(tensors_); }

  std::vector<TFE_TensorHandle*> tensors() const {

      parallel_device.Execute(context.get(), std::vector<ParallelTensor*>(),
                              "VarHandleOp", TFE_OpGetAttrs(handle_op.get()),
                              /*expected_max_outputs=*/1, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  const std::vector<std::unique_ptr<ParallelTensor>>& handles = *outputs;
  std::vector<ParallelTensor*> handle_inputs;

  int num_elems = TF_TensorElementCount(theta_tensor);
  vector<float> theta_data(num_elems);
  memcpy(theta_data.data(), TF_TensorData(theta_tensor),
         TF_TensorByteSize(theta_tensor));

  // Initialize space for the numerical gradient.
  vector<float> dtheta_approx(num_elems);

  // Get theta shape and store in theta_dims.
  int num_dims = TF_NumDims(theta_tensor);
  vector<int64_t> theta_dims(num_dims);

  // Process inputs.
  std::vector<tensorflow::OutputTensor> input_tensors;
  std::unordered_map<const Node*, std::vector<int>> input_nodes;
  status->status = tensorflow::ProcessInputs(fn_body, fn_name, ninputs, inputs,
                                             &input_tensors, &input_nodes);
  if (TF_GetCode(status) != TF_OK) return nullptr;

  // Process outputs.
  std::vector<tensorflow::OutputTensor> output_tensors;

};

absl::optional<std::vector<MaybeParallelTensorOwned>> ExecuteWithSpecialOps(
    const ParallelDevice& parallel_device,
    const std::string& parallel_device_name, TFE_Context* context,
    std::vector<MaybeParallelTensorUnowned> inputs, const char* operation_name,
    const TFE_OpAttrs* attributes, int expected_max_outputs,
    TF_Status* status) {
  absl::optional<std::vector<MaybeParallelTensorOwned>> result;

  if (!status->status.ok()) return;

  // Initialize a input_tensor vector with `nullptr` values.
  std::vector<const Tensor*> input_tensors_vector(num_inputs, nullptr);
  // A vector to keep track of newly created `tf::Tensor` objects.
  std::vector<Tensor> all_input_tensors;
  // Update the vector with information from `input_tensors` if provided.
  if (input_tensors != nullptr) {

  virtual bool UsesTFRT() = 0;

  // List attributes of available devices
  virtual void ListDevices(std::vector<DeviceAttributes>* devices) = 0;

  // Add `devices` into context's device manager. Context's device manager
  // will take ownership and maintain devices' lifetime.
  virtual absl::Status AddDevices(
      std::vector<std::unique_ptr<Device>> devices) = 0;

  // Block until all pending nodes are finished.

                           const string& op_name) {
  std::vector<int64_t> input_ids(inputs.size());
  std::vector<tensorflow::DataType> input_dtypes(inputs.size());
  for (int i = 0; i < inputs.size(); i++) {
    input_ids[i] = ToId(inputs[i]);
    input_dtypes[i] = inputs[i]->DataType();
  }
  std::vector<TapeTensor> tape_tensors;
  tape_tensors.reserve(outputs.size());
  for (auto t : outputs) {