Anyone had success in LORA training flux1? (not on an h100)

[pinballelectronica]

Curious what settings people have used to actually get a practical training speed on a consumer GPU. Floats, optimizers, ranks, alphas. My go-to is Prodigy but due to the ramp up time that may not be ideal. My 2x4090's are no match for float16 across the board. 95 steps a NIGHT at very conservative values :) I
hear nf4 produces very poor output so. Trying to find a sweet spot. Thank you.
This image took ~7 minutes at FP16.

[FurkanGozukara]

everything you say doesn't make sense with 2x4090 :D

even 8gb gpus are able to train FLUX dev model and starting from 16 GB you can train with best quality

rtx 4090 can train very fast and 2x trains even faster - lora

rtx 4090 can train a decent speed full fine tuning as well

[Seedmanc]

8Gb

Can't train even XL on 8 without going sub-1024 and 8 bit base model loading, I can't imagine what compromises must be taken to fit Flux there.