No, I mean factorizing the attention matrices themselves and replacing them with low-rank approximations after training.