Speculative Decoding on RL Rollout Prompts
167.6Rollout Time (s)BubbleSpec
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| BubbleSpecModel=Qwen3-1.7B2026.05 | 167.6 | 2,913 | 15,908 | 6,592 | 39,685 | 2.15 | 3.84 | 29.94 | |
| BubbleSpecModel=Qwen3-4B2026.05 | 235 | 2,895 | 17,616 | 6,499 | 38,577 | 2.09 | 3.81 | 28.6 | |
| VerlModel=Qwen3-1.7B2026.05 | 306.2 | 6,741 | 39,531 | 6,741 | 39,361 | — | — | — | |
| VerlModel=Qwen3-4B2026.05 | 399.6 | 6,489 | 37,910 | 6,481 | 37,910 | — | — | — | |
| BubbleSpecModel=Qwen2.5-VL-7B2026.05 | 400.1 | 4,853 | 34,027 | 9,414 | 59,912 | 1.81 | 3.73 | 21.7 | |
| VerlModel=Qwen2.5-VL-7B2026.05 | 571.9 | 9,500 | 60,509 | 9,500 | 60,509 | — | — | — |