performance-flash-attention-slid-when-using-flash-attention-with-sliding-window-and-cc1a0650

THEN Change the flash_kwargs initialization in modeling_flash_attention_utils.py from `{"window_size": (sliding_window, sliding_window)}` to `{"window_size": (sliding_window - 1, sliding_window)}` when sliding window is enabled. This ensures the flash attention implementation matches the behavior of other sliding window implementations combined with a causal mask.

flash_kwargs = {"window_size": (sliding_window - 1, sliding_window)} if use_sliding_windows else {}