I tried to quantize the model into BF16 and FP16 to preserve a bit more precision than NF4/FP8 while still running fast on data center cards. However, when I tried to hack the way to get it to run in ForgeUI, I can only get it to generate nonsense donut pictures (just like the 1st image of black hole).
Any ideas of where I need to change/hack to get it to work? Feels like there is a data scramble happening somewhere
@lllyasviel hopefully we can get this as a native feature : ]
I tried to quantize the model into BF16 and FP16 to preserve a bit more precision than NF4/FP8 while still running fast on data center cards. However, when I tried to hack the way to get it to run in ForgeUI, I can only get it to generate nonsense donut pictures (just like the 1st image of black hole).
Any ideas of where I need to change/hack to get it to work? Feels like there is a data scramble happening somewhere
@lllyasviel hopefully we can get this as a native feature : ]
Thanks XD K.