After updating the version, I attempted to train qwen2_vl but encountered issues with slower training speed and decreased accuracy. I have not been able to identify the cause.

[xueaa]

Reminder

• I have read the above rules and searched the existing issues.

Description

After updating the version, I attempted to train qwen2_vl but encountered issues with slower training speed and decreased accuracy. I have not been able to identify the cause.

Pull Request

No response

[xueaa]

When using the previous model for inference, I also encountered a significant drop in accuracy. I tried both the past version and the updated version, and the accuracy difference is more than 20%. Why do different versions cause such issues? Has anyone experienced similar situations?

[xueaa]

This is the current accuracy, but after switching to the previous version of LLaMA-Factory, the accuracy improved by about 20%. Why is there such a significant difference in performance even in deployment?

[hiyouga]

Is this commit the cause of this phenomenon? bc298c6

[xueaa]

这是造成这种现象的原因吗？BC298C6

Thank you for your response. I can adjust this and try, but I believe it's not related to this issue. The reason is that some bounding boxes (bbox) are correct, while others are incorrect, even though the input images are of the same size. This suggests that the problem might not be caused by the version conflicts we discussed.