Keep pulling the thread on Liang Wenfeng.
On offline benchmarks, DSpark substantially improves the accepted token length over state-of-the-art autoregressive and parallel drafters.
When deployed within the DeepSeek-V4 serving system under live user traffic, DSpark successfully mitigates verification waste.
Compared to the MTP-1 production baseline, DSpark accelerates per-user generation speeds by 60% to 85% at matched throughput levels.
DSpark enables previously unattainable performance tiers in serving systems by preventing severe throughput degradation under strict interactivity constraints.
Parallel drafters for Large Language Model inference suffer from rapid acceptance decay due to a lack of inter-token dependencies.
Indiscriminately verifying extended token blocks in parallel drafting wastes batch capacity and degrades throughput in high-concurrency serving systems.
DSpark is a speculative decoding framework that unifies high-throughput parallel generation with adaptive, load-aware verification.
DSpark utilizes a semi-autoregressive architecture, coupling a parallel backbone with a lightweight sequential module, to introduce intra-block dependency modeling and mitigate suffix decay.
DSpark employs confidence-scheduled verification, which dynamically tailors the verification length for each request based on estimated prefix survival probabilities and engine-specific throughput profiles.