“The 1/r scaling factor in the Low-rank adaptation (LoRA) parametrization makes the optimal learning rate approximately independent of the rank.”