CUDA_VISIBLE_DEVICES=0 deepspeed finetuning_pt.py --num_train_epochs 5 --train_batch_size 2 --pre_seq_len 16