SQD Scheme Optimizes Heterogeneous Disaggregation for Subquadratic Attention

The SQD scheme disaggregates inference by splitting decode into quadratic and subquadratic attention components rather than by operator type. This approach targets throughput and energy efficiency gains on heterogeneous DRAM and SRAM-based systems by accounting for the specific memory footprints of subquadratic variants.