FlexAttention for Efficient High-Resolution Vision-Language ModelsJunyan LiDelin Chenet al.2024ECCV 2024