export CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch train.py