Reproduce “When Parallel Drafter Meets Parallel Speculative Decoding” on OpenResearch

About this paper

Fuliang Liu, Xue Li, Kun Qian, Zhibin Wang, Wanchun Dou, Wenyuan Yu, Chen Tian

DPara overlaps speculative decoding’s parallel backbone with target verification by precomputing draft representations for every possible accepted-prefix boundary, then uses a lightweight autoregressive head with the revealed bonus token to continue drafting. In batch-size-1 greedy evaluations, it averages 3.21× and 3.52× speedups over autoregressive decoding on Qwen3-8B and Qwen3-14B, respectively, under the tested conditions; the head remains serial.

Key insights

• Computing branches for every acceptance boundary removes the specific prediction-induced fallback used by methods such as SSD, though the precomputation itself still incurs cost.
• Latency can matter more than acceptance length: DPara is faster than DSpark despite accepting about 4.4 tokens on average versus DSpark’s 5.36, because much of its backbone work is hidden during verification.
• Deployment trade-offs depend on hardware placement: a separate drafting GPU preserves most of DPara’s speedup, while single-GPU colocation loses its advantage over DSpark at larger batch sizes.

Summary by alphaXiv.