}
}

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;

    : public std::function<Status(const std::vector<Gradient*>&,
                                  std::vector<Gradient*>*, bool)> {
 public:
  template <typename lambda_type>
  explicit ForwardFunction(lambda_type lambda)
      : std::function<Status(const std::vector<Gradient*>&,
                             std::vector<Gradient*>*, bool)>(lambda) {}
};

// Computes Jacobian-vector products using forward-mode automatic

        }

        // reactor failure modes
        if (t instanceof RuntimeException || !(t instanceof Exception)) {
            // fail fast on RuntimeExceptions, Errors and "other" Throwables
            // assume these are system errors and further build is meaningless
            buildContext.getReactorBuildStatus().halt();

  TensorHandlePtr three_vector =
      VectorFloatTensorHandle({5., 6., 7.}, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  std::vector<TensorHandlePtr> vector_handles;
  vector_handles.reserve(2);
  vector_handles.push_back(std::move(two_vector));
  vector_handles.push_back(std::move(three_vector));
  std::unique_ptr<ParallelTensor> unknown_length_vector =

                                                           double growth_factor,
                                                           int bucket_count) {
  return new TFE_MonitoringBuckets([scale, growth_factor, bucket_count]() {
    return tensorflow::monitoring::Buckets::Exponential(scale, growth_factor,
                                                        bucket_count);
  });
}

  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.
  Status SummarizeValue(std::string& summary);

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

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

  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) {

  device->delete_device = &DeleteParallelDevice;
  device->execute = &ParallelDeviceExecute;
  std::vector<std::string> underlying_devices_vector;
  underlying_devices_vector.reserve(num_underlying_devices);
  for (int device_index = 0; device_index < num_underlying_devices;
       ++device_index) {
    underlying_devices_vector.push_back(underlying_devices[device_index]);
  }
  std::unique_ptr<ParallelDevice> parallel_device(

  // This cache has space for 4 blocks; we'll read from two files.
  const size_t n = 3;
  tf_gcs_filesystem::RamFileBlockCache cache(8, 32, 0, fetcher);
  std::vector<char> out;
  std::vector<char> a(n, 'a');
  std::vector<char> b(n, 'b');
  std::vector<char> A(n, 'A');
  std::vector<char> B(n, 'B');
  // Fill the cache.
  TF_EXPECT_OK(ReadCache(&cache, "a", 0, n, &out));
  EXPECT_EQ(out, a);
  EXPECT_EQ(calls, 1);

                           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) {