What is NOT used by ZeRO to boost memory efficiency for training large deep learning models? $($A$)$ It partitions model states and replicates them across data$-$ parallel processes; $($B$)$ It reduces activation memory; $($C$)$ It reduces the residual memory by temporary buffers and memory fragmentation; $($D$)$ It eliminates memory redundancies and makes the full aggregate memory capacity of a cluster available. Which of these choices is not used by ZERO $?$