// `tensor_id`, associated with `session`. There must be a graph node named
// "fifo_queue_enqueue_<tensor_id>", to be executed by this API call. It reads
// from a placeholder node "arg_tensor_enqueue_<tensor_id>".
//
// `tensor` is still owned by the caller. This call will be blocked if the queue
// has reached its capacity, and will be unblocked when the queued tensors again
// drop below the capacity due to dequeuing.
//

// have been executed. Note that "execution" here refers to kernel execution /
// scheduling of copies, etc. Similar to sync execution, it doesn't guarantee
// that lower level device queues (like GPU streams) have been flushed.
//
// This call may not block for execution of ops enqueued concurrently with this
// call.
TF_CAPI_EXPORT extern void TFE_ExecutorWaitForAllPendingNodes(

  // re-use a thread.
  std::vector<std::unique_ptr<DeviceThread>> device_threads_;
  // A cancellation manager to use if the caller does not provide one. When ops
  // are executed asynchronously this must outlive the queued op, so it can't be
  // function-local to Execute.
  mutable std::unique_ptr<CancellationManager> default_cancellation_manager_;
};

// Contains a tuple of tensors, one on each of the `underlying_devices_` of the