Bug related with your finetune script

[massyzs]

System Info / 系統信息

Hi, thanks for your wonderful work.
However, I found there might be a bug in your lora_trainer.py.
In function "compute_loss", you calculate mse between the velocity and clear latent feature, which might be wrong.
You may want MSE(noisy_latent-predicted_noise, latent) instead of MSE(v, latent).

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

Just run the finetuning code

Expected behavior / 期待表现

Should fix as target MSE(target_v,v) or MSE(predicted_latent,latent)