DSpark is a speculative decoding framework designed to accelerate Large Language Model (LLM) inference by combining semi-autoregressive draft generation with adaptive, load-aware verification. It addresses the acceptance decay and throughput degradation issues found in existing parallel drafters.
- DSpark utilizes a semi-autoregressive architecture that couples a parallel backbone with a lightweight sequential module to model intra-block dependencies.
- The system employs confidence-scheduled verification to dynamically tailor verification length based on prefix survival probabilities and throughput profiles.
- On offline benchmarks, DSpark substantially improves accepted length over state-of-the-art autoregressive and parallel drafters.
- In the DeepSeek-V4 serving system under live traffic, DSpark accelerates per-user generation speeds by 60 to 85 percent compared to the MTP-1 baseline.
By preventing verification waste and mitigating throughput degradation under strict interactivity constraints, DSpark enables performance tiers previously unattainable in production environments.