--optim 'adamw_torch_fused' \
    --seed 0 --data_seed 0 \
    --dataloader_num_workers 30 \
    --per_device_train_batch_size 2 --per_device_eval_batch_size 2 \
    --low_cpu_mem_usage false \
    --evaluation_strategy "steps" --eval_steps 5000 \
    --save_strategy "steps" --save_steps 5000 \
    --save_total_limit 2 \
    --gradient_accumulation_steps 4 \
    --learning_rate 3e-4 \
    --lr_scheduler_type "cosine" \
    --warmup_steps 28 \
    --num_train_epochs 5 \
    --use_lora \
    --lora_rank 64 \
    --lora_alpha 16 \
    --lora_dropout 0.05 \