希望能恢复lora权重文件能生成.bin格式，新版本是.safetensors，希望能加入参数控制，llama.cpp的qlora要bin格式

[lizhiling12345]

Reminder

• I have read the README and searched the existing issues.

Reproduction

python src/train_bash.py
--stage pt
--model_name_or_path /hy-tmp/model/AquilaCode-multi
--do_train
--dataset my_code
--template aquila
--lora_target all
--finetuning_type lora
--output_dir sft_checkpoint_aquila
--overwrite_cache
--per_device_train_batch_size 4
--gradient_accumulation_steps 1
--lr_scheduler_type cosine
--logging_steps 1
--save_steps 100
--save_total_limit 10
--learning_rate 5e-5
--num_train_epochs 1000
--plot_loss
--fp16
--quantization_bit 4

Expected behavior

No response

System Info

No response

Others

No response

[hiyouga]

--save_safetensors False