standard lora, α=r=32;

fp8 unet and full bf16 training;

using/trying prodigy plus optimizer.