performance-flash-attention-inte-qwen2-5-vlvisionattention-with-flash-attention-cra-3ee0408c

THEN Add the `is_causal` attribute to the `Qwen2_5_VLVisionAttention` class. Set it to `True` (or appropriate value) to match causal attention behavior required by flash attention. Alternatively, upgrade to the patched version (transformers v4.55+ or the commit containing PR #39121).

class Qwen2_5_VLVisionAttention(nn.Module):
    def __init__(self, ...):
        super().__init__()
        self.is_causal = True  # Add this line
        ...
    ...