DCOCO paper: contextual correction makes parallel speculative decoding up to 6.6x faster
burny_tech · x · 2026-10-30
A new arXiv paper introduces DCOCO (DFlash Contextual Corrector), fixing the core weakness of parallel speculative decoding. Parallel diffusion drafters propose multiple tokens at once but can't see which earlier tokens were selected, breaking intra-block dependencies. DCOCO keeps multiple candidate tokens per position and scores their contextual compatibility across adjacent draft positions in parallel, restoring dependencies at low cost. Across 4 verifier backbones and 10 benchmarks it achieves the highest average throughput in every configuration, up to 6.6x speedup over vanilla decoding and 6-23% higher throughput than the strongest competitor under 4-32 concurrent requests. Code and models are open source.
More from Infra
- AI-driven scientific discoveries cost $10M+ compute on Navier-Stokes, $100M+ overall — kimmonismus · 2026-10-30
- Google adds the most compute capacity by 2027 yet trades cheapest among hyperscalers — inductionheads · 2026-10-30
- Swapping 5600X for 5600G slashes Vulkan LLM inference 4x while ROCm stays unaffected — Jebbyk1 · 2026-10-30
- Situational Awareness is raising money for neocloud startup Fluidstack, months after near-collapse — nmasc_ · 2026-10-30
- 16GB GPU bandwidth-per-dollar tier list puts RX 6800 on top — HankYeomans · 2026-10-30
- First true omni model runs fully in-browser: 280 tok/s with WebGPU — victormustar · 2026-10-30