When extracting loras from models, use bf16 checkpoints
Subj. I noticed that you're extracting from krea2_turbo_fp8_scaled as a base model. This is wrong; even if the finetuned model you're processing is not available in bf16, you should always use bf16 for the base checkpoint (in this case, instead of the krea2_turbo_fp8_scaled). If you can obtain bf16 versions of the finetuned models - the better. Currently, you are just multiplying quantization errors unfortunately.
It may seem counterintuitive, but using bf16 (base) really helps even if your target (finetuned) model is a lower quant.
Agree that better to use 16 bit versions.
I reupload all loras. they work (problem was in naming of internal keys of layers). my script takes input two files base lora and fineruned. NOW any of inputs can be in ANY format (int8, canvrot, fp8, fp8, bf16, int4 convrot any known) the CONVERTED in memery to fp16 or bf16 , processed and writeted as you wish.
I agree, ok, but show me the error berween 16 and 8 with examples, its minirable :)
Big thank you! Downloading :)
Testing them now. Most of them work (with varying degree of success); gonzalomo currently broken (generates stripes and blocks of color).