All new LORAs produce "modules.devices.NansException: A tensor with all NaNs was produced in Unet" in StableDiffusion

Was using very old version, replaced with fresh install to be able to train SDXL and all new LORA become broken.
Reducing learning rate or training steps to very low allows to generate but only on level where model not actually trained yet and not giving anything close to dataset - so its useless anyway.

Start by reproducing SDXL LORA training from the fresh install and trace the reported modules.devices.NansException back to the failing training stage. Compare behavior with the older version and reduced learning-rate or step settings; done means newly trained LORAs no longer produce all-NaN tensors while learning the dataset.

Written by the indexing model from the issue text.