[arugulum]

>Then you'd have to compute the gradients for the whole network

You have to do that with LoRA regardless, to compute the gradients for the lowest-level LoRA weights.