// A non-blocking version of `Execute`. After each call, `Join` must be called
  // before `StartExecute` is called again. Using `StartExecute` with `Join`
  // allows the caller to schedule computation on multiple ParallelDevices
  // without sequencing those operations (first call `StartExecute` on each
  // parallel device, then call `Join` on each; even if some of the `Join`s

  // functions (and hence the tensors they keep alive). Instead, everything
  // is deleted in ~GradientTape. Persistent GradientTapes are useful when
  // users want to compute multiple gradients over the same tape.
  explicit GradientTape(bool persistent) : persistent_(persistent) {}
  ~GradientTape() {
    for (const auto& pair : op_tape_) {

  }
};

// 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)

template <std::size_t num_replicas>
TensorHandlePtr CreatePerDeviceValues(
    TFE_Context* context,
    const std::array<TFE_TensorHandle*, num_replicas>& components,
    const char* device, TF_Status* status);

TensorHandlePtr Multiply(TFE_Context* context, TFE_TensorHandle* first,
                         TFE_TensorHandle* second, TF_Status* status);

// Assert that `handle` is equal to `expected_value`.
template <typename value_type>

  void (*cleanup)(TF_Filesystem* filesystem);

  /// Creates a new random access read-only file from given `path`.
  ///
  /// After this call `file` may be concurrently accessed by multiple threads.
  ///
  /// Plugins:
  ///   * Must set `status` to `TF_OK` if `file` was updated.
  ///   * Must set `status` to `TF_NOT_FOUND` if `path` doesn't point to an