// call Session::Extend().
bool ExtendSessionGraphHelper(TF_Session* session, TF_Status* status) {
  if (session->graph != nullptr) {
    // Take the graph lock before the session lock to avoid deadlock. This is
    // safe since session->graph does not change.
    session->graph->mu.lock();
    mutex_lock session_lock(session->mu);
    const Graph& graph = session->graph->graph;

  // threadpool, so that we avoid the possibility of running the runner_ in the
  // threadpool of GPU event mgr, as that can trigger more callbacks to be
  // scheduled on that same threadpool, causing a deadlock in cases where the
  // 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).

                        const int in_flight_nodes_limit)
      : status_(TF_NewStatus()),
        // If the context's default exector is set to async, re-using that in
        // each thread would cause collectives to deadlock. For consistency we
        // create a new sync executor for every thread.
        //
        // TODO(allenl): We should have an async API that works with the
        // parallel device.
        device_(device),

  // parallel device, then call `Join` on each; even if some of the `Join`s
  // return a bad status the caller must run all of the `Join`s or any future
  // `StartExecute`s will deadlock).
  //
  // If `is_async=false` (constructor argument), `cancellation_manager` must
  // live until `Join` finishes. If `is_async=true` it must live until `Join` is