Creates lora from two checkpoints by extracting the difference.
model_a is the base model
model_b is the finetuned model
rank is the final rank of the model, higher = larger size
method is rsvd (fast, but marginally less precise) or svd
There are two ways to extract, rsvd which is faster and 99% as good as svd, you can choose oversampling for this one.