Add siglip flashattention support?

[lucasjinreal]

Add siglip flashattention support?

I noticed that the navit-flashatten-siglip version already added flash attention, would transformers add it support inside ?

[LysandreJik]

cc @molbap or @qubvel maybe :)

[lucasjinreal]

these two guys looked but didn't response.

[qubvel]

@lucasjinreal thanks for the feature request, I will take a look at it!

[lucasjinreal]

thanks, since more and more MLLM used siglip as vision encoder, if flash attention support would be very reduce training cost time