Single SAEs trained on the residual stream activation vectors from every transformer layer simultaneously (including the transformers).