// caller of event_mgr->ThenExecute() blocks on the completion of the callback
  // (as in the case of ConstOp kernel creation on GPU, which involves copying a
  // CPU tensor to GPU).
  // Setting a larger thread pool does not help with the Swift caller, as we use
  // a different TFE context for each thread of execution (for running graph
  // functions, and their send/recvs corountines).
  config.set_inter_op_parallelism_threads(1);

  //   Outputs:
  std::vector<TensorHandlePtr> op_outputs_ TF_GUARDED_BY(execution_mutex_);
  // TF_Status is an incomplete type and so can't be stack allocated. To avoid
  // unnecessary allocations each Execute call, we keep one heap-allocated
  // version for the thread.
  StatusPtr status_ TF_GUARDED_BY(execution_mutex_);

  const std::string device_;
  ExecutorPtr executor_ TF_GUARDED_BY(execution_mutex_);