How to choose distributed training mode

As far as I know, TensorFlow models only support PS-Worker mode and XGBoost models only support allreduce mode to train. Can we support nccl/mpi allreduce mode for TensorFlow? Allreduce may be quite important especially in GPU training.