MLPerf Training introduces DLRMv4, a next-generation recommendation benchmark built on HSTU sequence modeling and the Yambda-5B dataset with a 560 GB production-class embedding footprint.- Focus keyphrase: DLRMv4 MLPerf Training