Under this pruning scheme, the token-level latency reduction translates into roughly 78% end-to-end speedup under typical request patterns.