ASPIRE combines a unified target-model forward that mixes per-request sparse-attention drafting and full-attention verification, an online per-request scheduler based on smoothed acceptance and a batch-aware draft/verify cost ratio, and one full-attention refresh layer during each draft step to update sparse KV pages. · CiteArk
ASPIRE combines a unified target-model forward that mixes per-request sparse-attention drafting and full-attention verification, an online per-request scheduler based on smoothed acceptance and a batch-aware draft/verify cost ratio, and one full-attention refresh layer during each draft step to update sparse KV pages.
