Source weights?
Hey,
I'm doing some quants for comfy-org repo and seeing how community quants are being done, what did you use for source on this one?
Asking because however I do it, it ends up close to silver's quant, so only explanation I can think of for your measurements to show it as worse is that you used the comfy-org model as source, which sadly are extra lossy since comfy doesn't have fp8 rowwise support, so the fp8 scaled is tensorwise requant of that, this would explain those numbers.
The proper way for Ideogram4 would be to requant the original fp8 rowwise.
Yep, I sure stepped right in that trap. I based these on said comfy version.
Yep, I sure stepped right in that trap. I based these on said comfy version.
Thanks for confirming, and apologies for that trap, the whole model release was rushed mess as we didn't get access to the full weights either and had to rely on their quants.