DSpark is a speculative decoding framework designed for high-concurrency LLM serving. It combines a parallel backbone with a lightweight sequential module to model dependencies within drafted blocks and reduce suffix acceptance decay. Its confidence-scheduled verification policy adapts verification length using estimated prefix survival probabilities and engine-specific throughput profiles. The paper reports higher accepted lengths than autoregressive and parallel drafters on offline benchmarks. Under live traffic in a DeepSeek-V4 serving system, DSpark reportedly improved per-user generation speed by 60% to 85% over the MTP-1 production baseline at matched throughput, while avoiding severe degradation under strict interactivity constraints.
No heat snapshots are available in the last 24 hours.