std::vector<TFE_TensorHandle*> tensors() const {
    std::vector<TFE_TensorHandle*> result;
    result.reserve(tensors_.size());
    for (const TensorHandlePtr& tensor : tensors_) {
      result.emplace_back(tensor.get());
    }
    return result;
  }

 private:
  ParallelTensor(const ParallelDevice& device,
                 std::vector<TensorHandlePtr> tensors,

TF_CAPI_EXPORT extern void TF_DeleteFile(const char* filename,
                                         TF_Status* status);

// Retrieves the next item from the given TF_StringStream and places a pointer
// to it in *result. If no more items are in the list, *result is set to NULL
// and false is returned.
//
// Ownership of the items retrieved with this function remains with the library.

namespace tensorflow {
namespace parallel_device {

// Allocate a parallel device named `device_name` which forwards operations to
// `underlying_devices`, maintaining "parallel tensors" with components placed
// on each underlying device.
//
// For example if `device_name` is
//   "/job:localhost/replica:0/task:0/device:CUSTOM:0"
// and `underlying_devices` is
//   {"/job:localhost/replica:0/task:0/device:GPU:0",

// loading a library. The rules for determining the exact location of the
// library are platform-specific and are not documented here.
//
// On success, place OK in status and return the newly created library handle.
// The caller owns the library handle.
//
// On failure, place an error status in status and return NULL.
TF_CAPI_EXPORT extern TF_Library* TF_LoadLibrary(const char* library_filename,

void RegisterParallelDevice(
    TFE_Context* context, const char* device_name,
    const std::array<const char*, num_devices>& underlying_devices,
    TF_Status* status);

// Create and modify a variable placed on a parallel device which composes
// `first_device` and `second_device`.
void BasicTestsForTwoDevices(TFE_Context* context, const char* first_device,
                             const char* second_device);

  std::vector<const char*> raw_strs;
  std::vector<size_t> lengths;
  raw_strs.reserve(values.size());
  lengths.reserve(values.size());
  for (const auto& s : values) {
    raw_strs.emplace_back(s.data());
    lengths.emplace_back(s.size());
  }
  return SetAttrStringList(attr_name,
                           reinterpret_cast<const void**>(raw_strs.data()),
                           lengths.data(), values.size());
}

  template <typename G>
  Cleanup(Cleanup<G>&& src)  // NOLINT
      : released_(src.is_released()), f_(src.release()) {}

  // Assignment to a Cleanup object behaves like destroying it
  // and making a new one in its place, analogous to unique_ptr
  // semantics.
  Cleanup& operator=(Cleanup&& src) {  // NOLINT
    if (!released_) f_();
    released_ = src.released_;
    f_ = src.release();
    return *this;
  }

  ~Cleanup() {

// Create a new TFE_TensorHandle with the same contents as 'h' but placed
// in the memory of the device name 'device_name'.
// If source and destination are the same device, then this creates a new handle
// that shares the underlying buffer. Otherwise, it currently requires at least
// one of the source or destination devices to be CPU (i.e., for the source or
// destination tensor to be placed in host memory).

  // ForwardAccumulator.
  explicit ForwardAccumulator(
      const VSpace<Gradient, BackwardFunction, TapeTensor>& vspace,
      bool use_batch)
      : vspace_(vspace), use_batch_(use_batch) {
    call_state_.emplace(nullptr, false);
  }

  virtual ~ForwardAccumulator() {
    for (auto accumulated : accumulated_gradients_) {
      vspace_.DeleteGradient(accumulated.second);
    }
  }

// library are platform-specific and are not documented here.
//
// On success, returns the newly created library handle and places OK in status.
// The caller owns the library handle.
//
// On failure, returns nullptr and places an error status in status.
TF_CAPI_EXPORT extern TF_Library* TF_LoadPluggableDeviceLibrary(
    const char* library_filename, TF_Status* status);