Any way to reduce prodigy memory usage for SDXL finetune?

[ppbrown]

I have a 4090.
I'm wondering if there is ANY way to reduce prodigy memory usage so that it will fit in 24gigs, to do a finetune for sdxl?

I thought I'd try the "reduce unet to float8" thing, but then python throws some kind of error like

foreach_tensor_norm_cuda not implemented for float8

[mx]

Unfortunately, not really. Prodigy is just memory hungry.