//
  // TFE_OpGetDevice(op) records the original placement of the operation. It may
  // be an empty string if no device was explicitly requested, but will
  // otherwise be the name of this custom device. Ops are placed onto a custom
  // 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,

  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,

  // 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);

// 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).