Need Option to Disable Flash Attention in VideoLLaMA2.1-7B-AV (SiglipVisionModel)

[harshmoothat]

When using SiglipVisionModel inside VideoLLaMA2.1-7B-AV, I encounter the following error:

ValueError: SiglipVisionModel does not support Flash Attention 2.0 yet.

I do not need Flash Attention for my use case and would like to disable it.
Could you provide an official way to toggle it off?

[zucchini-nlp]

Hmm Siglip does support FA2 as per this code block, Ca you check transformers version, and that FA2 was available for Siglip in that version? If not update to the latest

https://github.com/huggingface/transformers/blob/main/src/transformers/models/siglip/modeling_siglip.py#L439