是否已有关于该错误的issue或讨论？ | Is there an existing issue / discussion for this?

• 我已经搜索过已有的issues和讨论 | I have searched the existing issues / discussions

该问题是否在FAQ中有解答？ | Is there an existing answer for this in FAQ?

• 我已经搜索过FAQ | I have searched FAQ

当前行为 | Current Behavior

Over last 3 months, I've been working on the Qwen2.5-VL from huggingface's transformers.

There is a persistent performance drop in it due to the features introduced in huggingface. The causes are wrong 3D rope PE. I have to make a lots of monkey patch to make sure they are correct.....

I wonder are you aware of this? Can you work with huggingface to solve the problem?

I can't even image how much impact this will be if tons of people are using a wrong Qwen2.5-VL implementation..

Issues raised by users:

1. Qwen2.5VL is broken! huggingface/transformers#40154
2. Qwen2.5-VL-7B-Instruct: Significant accuracy regression on MMMU benchmark with transformers >=4.54.0 huggingface/transformers#40136
3. Qwen2.5-VL-7B-Instruct Accuracy Regression Still Persists in v4.56.2 huggingface/transformers#41180

PR attempt to fix: huggingface/transformers#40490

But seems like according to the latest issue huggingface/transformers#41180 the performance drop still exist.

期望行为 | Expected Behavior

Qwen2.5-VL should behave normally in inference...

复现方法 | Steps To Reproduce

No response

运行环境 | Environment

- OS:
- Python:
- Transformers:
- PyTorch:
- CUDA (`python -c 'import torch; print(torch.version.cuda)'`):

备注 | Anything else?

No response