status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  TensorHandlePtr value_one(FloatTensorHandle(3., status.get()));
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TensorHandlePtr value_two(FloatTensorHandle(-2., status.get()));
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  // Create a tensor on the first parallel device
  TensorHandlePtr value_one(FloatTensorHandle(1., status.get()));
  TensorHandlePtr value_two(FloatTensorHandle(2., status.get()));
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  std::array<TFE_TensorHandle*, 2> components{value_one.get(), value_two.get()};
  TensorHandlePtr first_combined_value = CreatePerDeviceValues(

  {
    TensorHandlePtr value_one(FloatTensorHandle(3., status.get()));
    TensorHandlePtr value_two(FloatTensorHandle(-2., status.get()));
    std::array<TFE_TensorHandle*, 2> components{value_one.get(),
                                                value_two.get()};
    TensorHandlePtr combined_value =

  TensorHandlePtr two_vector = VectorFloatTensorHandle({3., 4.}, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  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);

  // dtype of read operations).
  TF_DataType type_;
};

// Creates a TFE_TensorHandle with value `v`.
TensorHandlePtr FloatTensorHandle(float v, TF_Status* status);

// Creates a rank-one TFE_TensorHandle with value `v`.
TensorHandlePtr VectorFloatTensorHandle(const std::vector<float>& v,
                                        TF_Status* status);

#include "tensorflow/core/framework/tensor_shape.h"
#include "tensorflow/core/framework/types.h"

namespace tensorflow {
namespace parallel_device {

using TensorHandlePtr = tensorflow::Safe_TFE_TensorHandlePtr;

class ParallelTensor;
class DeviceThread;

// Forwards operations to `devices`, maintaining ParallelTensor with components
// placed on each underlying device.
class ParallelDevice {

    cancellation_manager_ = &cancellation_manager;
    execution_state_ = ExecutionState::kReadyToExecute;
  }
  start_execute_.notify_one();
}

std::vector<TensorHandlePtr> DeviceThread::Join(TF_Status* status) {
  std::vector<TensorHandlePtr> result;
  {
    tensorflow::mutex_lock l(execution_mutex_);
    while (execution_state_ != ExecutionState::kHasResult) {
      finished_execute_.wait(l);
    }

    return;
  }

  for (int i = 0; i < typed_outputs.size(); ++i) {
    MaybeParallelTensorOwned typed_output(std::move(typed_outputs[i]));
    if (absl::holds_alternative<TensorHandlePtr>(typed_output)) {
      outputs[i] = absl::get<TensorHandlePtr>(typed_output).release();
    } else {
      outputs[i] = ParallelTensorToTensorHandle(
                       named_device->name(), context,