Merge LoRA Adapters into Base Weights

A LoRA adapter stores a low-rank update $\Delta W = \frac{\alpha}{r} B A$ next to a frozen weight $W$ (shape (d_out, d_in)), with $A$ of shape (r, d_in) and $B$ of shape (d_out, r). For deployment you can merge it into the weight, so inference costs nothing extra: