deepspeed --hostfile hosts --num_gpus=4 src/train_bash.py \
    --stage sft \
    --do_train \
    --template llama2 \
    --dataset alpaca_gpt4_en,alpaca_gpt4_zh \
    --finetuning_type full \
    --model_name_or_path /work/home/liangjing/.cache/modelscope/hub/skyline2006/llama-7b \
    --output_dir /work/share/huchen1/liangjj/llama_factory \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 1 \
    --preprocessing_num_workers 2 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 100 \
    --eval_steps 100 \
    --learning_rate 5e-5 \
    --max_grad_norm 0.5 \
    --num_train_epochs 4.0 \
    --val_size 0.01 \
    --evaluation_strategy steps \
    --load_best_model_at_end \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --plot_loss \
    --fp16 \
    --save_on_each_node \
    --deepspeed deepspeed.json

#--resume_from_checkpoint /work/share/huchen1/liangjj/round6/round7/checkpoint_2000