“Block AttnRes is a variant of Attention Residuals that partitions layers into blocks and attends over block-level representations to reduce memory and communication overhead in large-scale model training.”