#deepspeech zero3 启动
torchrun --nproc_per_node 8 sft_train.py \
    --ddp_timeout 36000 \
    --model_name_or_path /path_to_llm/hf_llama_7b \
    --llama \
    --use_lora \
    --deepspeed configs/deepspeed_config_stage3.json \
    --lora_config configs/lora_config_llama.json \
    --train_file belleMath.json \
    --validation_file belleMath-dev1K.json \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 1 \
    --num_train_epochs 10 \
    --model_max_length 1024 \
    --save_strategy "steps" \
    --save_total_limit 3 \
    --learning_rate 3e-4 \
    --weight_decay 0.00001 \
    --warmup_ratio 0.01 \
    --lr_scheduler_type "cosine" \
    --logging_steps 10 \
    --evaluation_strategy "steps" \
    --torch_dtype "bfloat16" \
    --bf16 \
    --seed 1234 \
    --gradient_checkpointing \
    --cache_dir hf_cache_dir \
    --output_dir "exp" 
