[Question]: 训练LORA后模型输出错误

[NoahRe1]

Description

使用与Aquila-chat-lora.yaml的默认配置训练LORA，然后使用generate_chat_lora.py推理，模型输出为乱码：

(nndl) root@dl-2306261943226qo-pod-jupyter-7ccbcb95d8-sxjp9:~/FlagAI/examples/Aquila/Aquila-chat# python generate_chat_lora.py 
******************** lm aquila-7b
***************use cache True
model checkpoint_path=./checkpoints_in/aquila-7b/pytorch_model.bin are loaded successfully...
--------------------------------------------------------------------------------
text is Find the product of the numbers: 5 and 8
pred is  TheAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
--------------------------------------------------------------------------------
text is Provide five tips for effectively using tape measures
pred is  1AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
--------------------------------------------------------------------------------
text is Create a resume for a job in web development.
pred is  [AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

一些尝试：

1. 换用了若干lr值（2e-4、2e-6、2e-8甚至0），均会产生类似于上面的输出。
2. 不管是用自己的数据集，还是用sft_samples.jsonl和convo_samples.jsonl训练，均会产生类似于上面的输出。

Alternatives

No response

[Anhforth]

pull最新版本微调并推理，我加了.enable_input_require_grads()之后效果好多了

[BAAI-OpenPlatform]

此问题已关闭，如果还有问题可以重新打开