sft lora finetune的loss和gradient norm的数值过于大，sft训练后性能反而显著下降

我用finetune的代码在我的数据集上进行lora微调，finetune.py没有任何修改，但是训练过程中的loss和gradient norm的数值都是好几十，感觉不正常啊，训练完后性能也下降了，请问这是正常的吗？
我的训练Log数值：
{'loss': 17.0996, 'grad_norm': 14.1919527053833, 'learning_rate': 4.9146347821251266e-05, 'epoch': 0.55}
我的微调脚本：
accelerate launch --num_processes 4 finetune/finetune.py
--model_dir openmoss/MOSS-Audio-8B-Thinking
--data_path path/to/data
--eval_data_path path/to/data
--eval_strategy steps
--save_strategy steps
--output_dir path/to/output
--attn_implementation sdpa
--max_len 2048
--use_lora
--lora_rank 8
--lora_alpha 32
--lora_on_audio_encoder true
--bf16
--per_device_train_batch_size 1
--per_device_eval_batch_size 1
--gradient_accumulation_steps 8
--num_train_epochs 1
--learning_rate 1e-4
--logging_steps 1
--save_steps 10
--eval_steps 10
--lr_scheduler_type cosine
--warmup_steps 50
--save_total_limit 5
--max_grad_norm 1.0 \

Start by inspecting finetune/finetune.py and reproducing the reported run with the supplied accelerate command and training arguments. Compare the logged loss and grad_norm with the configured max_grad_norm, LoRA settings, and evaluation results; done means determining whether the behavior is expected or identifying a reproducible training or evaluation defect.