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,
                 absl::Span<const int64_t> shape, const TF_DataType dtype)

      });
}
bool Tape::ShouldRecord(
    absl::Span<const AbstractTensorHandle* const> tensors) const {
  std::vector<int64_t> tensor_ids(tensors.size());
  std::vector<tensorflow::DataType> tensor_dtypes(tensors.size());
  for (int i = 0; i < tensors.size(); i++) {
    tensor_ids[i] = ToId(tensors[i]);
    tensor_dtypes[i] = tensors[i]->DataType();
  }
  return GradientTape::ShouldRecord(tensor_ids, tensor_dtypes);
}

                       const string& op_name = "");
  // Returns whether any tensor in a list of tensors is being watched and has
  // a trainable dtype.
  bool ShouldRecord(
      absl::Span<const AbstractTensorHandle* const> tensors) const;
  // Unwatches this tensor on the tape. Mainly used for cleanup when deleting
  // eager tensors.
  void DeleteTrace(const AbstractTensorHandle*);

TF_AbstractTensor* TF_OutputListGet(TF_OutputList* o, int i);
// Append a tensor at the end of the output list, growing its size by one.
void TF_OutputListPushBack(TF_OutputList* o, TF_AbstractTensor* tensor,
                           TF_Status*);

// TF_ExecuteOperation will, if in eager mode, execute, if in graph mode, maybe
// capture some inputs and then add a node in the graph. The output tensors are
// returned through the provided TF_OutputList.

                                       TFE_TensorHandle* tensor,
                                       TF_Status* status, void* device_info) {
  TF_SetStatus(
      status, TF_UNIMPLEMENTED,
      absl::StrCat("Trying to copy a tensor ",
                   tensorflow::unwrap(tensor)->DebugString(),
                   " on to a parallel device. Pack non-parallel "
                   "tensors for each device into a parallel tensor explicitly.")

// an error.
//
// All component tensors must have the same dtype. Currently they must also have
// the same shape, although this requirement may be relaxed in the future.
//
// `device_name` must not name an existing physical or custom device (see
// the documentation for TFE_RegisterCustomDevice for more information).
//
// Tensors may be copied on or off the device explicitly using

// some input tensors are not on that device.
// LINT.IfChange
// Note: Keep in sync with internal copy of enum in eager/context.h.
typedef enum TFE_ContextDevicePlacementPolicy {
  // Running operations with input tensors on the wrong device will fail.
  TFE_DEVICE_PLACEMENT_EXPLICIT = 0,
  // Copy the tensor to the right device but log a warning.
  TFE_DEVICE_PLACEMENT_WARN = 1,

  for (int component_index = 0; component_index < tensors_.size();
       ++component_index) {
    // TODO(allenl): Add a C API for summarizing tensors. Currently custom
    // devices limiting themselves to a C API (for ABI compatibility) would need
    // to implement summarization for component tensors themselves.
    ImmediateExecutionTensorHandle* component =
        tensorflow::unwrap(tensors_[component_index].get());

// drop below the capacity due to dequeuing.
//
// Tensors are dequeued via the corresponding TF dequeue op.
// TODO(hongm): Add support for `timeout_ms`.
TF_CAPI_EXPORT extern void TF_EnqueueNamedTensor(TF_Session* session,
                                                 int tensor_id,
                                                 TF_Tensor* tensor,

  }
};

// An abstract operation describes an operation by its type, name, and
// attributes. It can be "executed" by the context with some input tensors.
// It is allowed to reusing the same abstract operation for multiple execution
// on a given context, with the same or different input tensors.
class TracingOperation : public AbstractOperation {
 protected:
  explicit TracingOperation(AbstractOperationKind kind)