transformers

Fix SDPA attention precision issue in Qwen2.5-VL

#37363

Merged