Jul 2026
DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation
DSpark is introduced, a speculative decoding framework that unifies high-throughput parallel generation with adaptive, load-aware verification, and enables performance tiers that were previously unattainable, shifting the Pareto frontier of the DeepSeek-V4 serving system.
Xin Cheng, Xingkai Yu, Chenze Shao et al.
· arXiv.org · 24 citations
· ⚡8