public:
  // Eager async execution is only supported when remote eager is not in use
  // (b/157523095).
  explicit ParallelDevice(const std::vector<std::string>& devices,
                          bool is_async = false, int in_flight_nodes_limit = 0);

  ~ParallelDevice();

  // Helper to copy a tensor handle from another device once for each component
  // of the ParallelDevice.
  //

                                         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.
// Item points are invalidated after a call to TF_StringStreamDone.

// names will be modified if their name already exists in the graph. If false,
// conflicting names will be treated as an error. Note that this option has no
// effect if a prefix is set, since the prefix will guarantee all names are
// unique. Defaults to false.
TF_CAPI_EXPORT extern void TF_ImportGraphDefOptionsSetUniquifyNames(
    TF_ImportGraphDefOptions* opts, unsigned char uniquify_names);

                                      bool reset_context, int keep_alive_secs,
                                      int64_t init_timeout_in_ms, int retries,
                                      bool clear_existing_contexts = false) = 0;

  // Initializes context for the local worker and no contexts will be created
  // for remote workers. Currently this only works for resetting context.

    released_ = true;
    return std::move(f_);
  }

  bool is_released() const { return released_; }

 private:
  static_assert(!std::is_reference<F>::value, "F must not be a reference");

  bool released_ = false;
  F f_;
};

template <int&... ExplicitParameterBarrier, typename F,
          typename DecayF = typename std::decay<F>::type>
Cleanup<DecayF> MakeCleanup(F&& f) {
  return Cleanup<DecayF>(std::forward<F>(f));

  // true if async execution is enabled.
  bool async = false;
  TFE_ContextDevicePlacementPolicy device_placement_policy{
      TFE_DEVICE_PLACEMENT_SILENT};
  // If true, use TFRT backend
  bool use_tfrt = false;
  // Whether to run elementary eager ops wrapped in a call op.
  bool run_eager_op_as_function = false;
  // Whether to rewrite jit_compile functions.
  bool jit_compile_rewrite = false;
};

  }
  if (call_state_.top().accumulating) {
    return false;
  }
  for (int i = 0; i < tensor_ids.size(); ++i) {
    if (accumulated_gradients_.find(tensor_ids[i]) !=
        accumulated_gradients_.end()) {
      if (IsDtypeTrainable(dtypes[i])) {
        return true;
      }
    }
  }
  return false;
}

template <typename Gradient, typename BackwardFunction, typename TapeTensor>

TF_CAPI_EXPORT unsigned char TF_GetXlaAutoJitEnabled();

// Sets XLA's minimum cluster size. This has global effect.
TF_CAPI_EXPORT void TF_SetXlaMinClusterSize(int size);

// Gets/Sets TF/XLA flag for whether(true) or not(false) to disable constant
// folding. This is for testing to ensure that XLA is being tested rather than
// Tensorflow's CPU implementation through constant folding.
TF_CAPI_EXPORT unsigned char TF_GetXlaConstantFoldingDisabled();

      ABSL_EXCLUSIVE_LOCKS_REQUIRED(mu_) {
    auto it = cache_.find(key);
    if (it == cache_.end()) {
      return false;
    }
    lru_list_.erase(it->second.lru_iterator);
    if (timer_seconds_() - it->second.timestamp > max_age_) {
      cache_.erase(it);
      return false;
    }
    *value = it->second.value;
    lru_list_.push_front(it->first);
    it->second.lru_iterator = lru_list_.begin();

 *
 * I.e. if y = <output of the forward model> and w = inputs[input_index],
 * this will calculate dy/dw numerically.
 *
 * `use_function` indicates whether to use graph mode(true) or eager(false).
 *
 * `numerical_grad` is the pointer to the AbstractTensorHandle* which will
 * hold the numerical gradient data at the end of the function.
 */
Status CalcNumericalGrad(AbstractContext* ctx, Model forward,