Load LoRA in cli_demo.py doesn't work

[Vicky0522]

System Info / 系統信息

I used the train_ddp_i2v.sh script to finetune LoRA, then loaded the saved LoRA weights with the --load_lora argument in cli_demo.py. However, the outputs look exactly the same as those produced without LoRA, despite seeing correct results during validation in the training process. It appears that the --load_lora option isn’t functioning as intended.

Information / 问题信息

• The official example scripts / 官方的示例脚本
• My own modified scripts / 我自己修改的脚本和任务

Reproduction / 复现过程

1. use train_ddp_i2v.sh script to finetune LoRA
2. load LoRA via cli_demo.py

Expected behavior / 期待表现

The results should be the same as the results during validation in the training process.

[holma91]

Thought I was going insane when I saw such a difference between wandb and locally! Can confirm that this is a bug for me aswell.

[zRzRzRzRzRzRzR]

We are reviewing this pull request and will respond soon.

[JuheonYi]

Hi, were you able to fully reproduce the validation results with the fix? Even after applying the scale correctly, there are still some difference...