// Returns the operation's device name.
  //
  // The value returned may be different from the one set by SetDeviceName, but
  // it will be compatible with it: the name will be updated by device placement
  // logic to refer to the specific device chosen.
  //
  // Example: If one calls `op->SetDeviceName("/device:GPU")`, the value
  // returned by DeviceName should be "/device:GPU:*" until a particular GPU is

  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;
  }

// 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).
// If async execution is enabled, the copy may be enqueued and the call will

int TF_OutputListNumOutputs(TF_OutputList* o);
// Return the `i`th output in the list.
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

    // Note: we hold onto mu_ for the rest of this function. In practice, this
    // is okay, as stat requests are typically fast, and concurrent requests are
    // often for the same file. Future work can split this up into one lock per
    // key if this proves to be a significant performance bottleneck.
    absl::MutexLock lock(&mu_);
    if (LookupLocked(key, value)) {
      return TF_SetStatus(status, TF_OK, "");
    }

    TF_Status* status);

// -----------------------------------------------------------------------------
// Eager Executor APIs.
typedef struct TFE_Executor TFE_Executor;

// Creates a new eager Executor. Nodes in one executor are guaranteed to be
// executed in sequence. Assigning nodes to different executors allows executing
// nodes in parallel.
// in_flight_nodes_limit: when is_async is true, this value controls the

#define TENSORFLOW_C_EAGER_C_API_REMOTE_TEST_UTIL_H_

// Run a function containing a MatMul op and check its output.
// If heavy_load_on_streaming_rpc is true, send some rpc requests before the one
// which creates a remote input, to simulate a scenario that the remote input
// is not ready when we start running an op or a function.
void TestRemoteExecuteSilentCopies(bool async, bool remote, bool func,

  std::pair<std::unique_ptr<TensorSliceReader::VarToShapeMap>,
            std::unique_ptr<TensorSliceReader::VarToDataTypeMap> >
  BuildV2VarMaps();

  // Invariant: exactly one of "reader_" and "v2_reader_" is non-null.
  std::unique_ptr<TensorSliceReader> reader_;
  std::unique_ptr<BundleReader> v2_reader_;

  std::unique_ptr<TensorSliceReader::VarToShapeMap> var_to_shape_map_;

// TF_TriState is the C API typedef for tri-state.
typedef enum TF_TriState {
  TF_TriState_Default = 0,
  TF_TriState_Off,
  TF_TriState_On,
} TF_TriState;

// TF_GrapplerItem represents a combination of a graph, one of more fetch nodes,
// and potentially a set of nodes to feed.
typedef struct TF_GrapplerItem TF_GrapplerItem;

// Flags indicating whether existing optimizers should be turned off.

  const std::vector<std::string> underlying_devices_;
  // A sequence of thread wrappers, one per device, for executing operations in
  // parallel.
  //
  // Conceptually this is a thread pool with one thread per device. It requires
  // less synchronization than a thread pool would for this task, since Execute
  // acquires each thread in order (and so only one Execute will schedule
  // blocking collective operations at a time), and avoids some dynamic