Figure 5 · DSpark speculative decoding

Adaptive DSpark keeps its lead at high concurrency

Muse-Spark-Open-816B-A42B-NVFP4 on 4 × NVIDIA GB300 (TP4), SPEED-Bench. One point per concurrency level, c1 to c256; up and to the right is better.

Per-user speed = 1 / median TPOT. Values read off the draft chart (about ±2%); replace with the measured table before publishing.