MASTER_PORT=$(shuf -n 1 -i 10000-65535)
HIP_VISIBLE_DEVICES=2 deepspeed --num_gpus=1 --master_port $MASTER_PORT src/train_bash.py \
    --stage sft \
    --do_train True \
    --flash_attn \
    --template llama2 \
    --dataset alpaca_gpt4_zh \
    --finetuning_type lora \
    --model_name_or_path /home/llama-2-13B \
    --lora_target q_proj,v_proj \
    --output_dir ./checkpoint_13b \
    --overwrite_output_dir \
    --per_device_train_batch_size 16 \
    --lr_scheduler_type cosine \
    --logging_steps 1 \
    --save_steps 1000 \
    --learning_rate te-5 \
    --gradient_accumulation_steps 1 \
    --bf16 \
    --deepspeed bf16_deepspeed.json \