It's worse than that, because lora requires two matrices per layer. At full rank, you have an additional NxN parameters to learn versus full finetuning, where N

It's worse than that, because lora requires two matrices per layer. At full rank, you have an additional NxN parameters to learn versus full finetuning, where N is min(input_features, output_features). For example, tuning a layer of 128 in x 256 out is 32k params. Learning a full-rank lora for that layer would be two matrices of 128x128 and 128x256 = 48k params.