pending_call_nodes;
  // Cache the host device for each TPU cluster.
  std::unordered_map<Operation*, std::optional<std::string>> cluster_to_host;

  auto insert_pending_op = [&](Operation* op,
                               tf_device::ClusterOp tpu_cluster) {
    // Add callee nodes to pending_call_nodes.
    if (CallOpInterface call = dyn_cast<CallOpInterface>(op)) {