SigLIP implementation

[alexisdrakopoulos]

Hi,

I just came across your paper, it seems that this also applies to SigLIP? Do you think it's worth trying a Long-SigLIP implementation?

[Larry-Gan]

long-siglip would be dope, but at this point you'll probably need to train it up yourself.
My understanding is that if you still stretch to 4, the effective context length would only be 196 though, so we may need to stretch to 5 to get 240