transformers

[qwen2-vl] fix vision attention scaling

#39043

Merged