TopK SAEs trained on the residual stream activation vectors from a single transformer layer, including the transformers.