DeepSeek开源DSpark推测解码框架，加速V4生成60-85%
DeepSeek Releases DSpark, a Speculative Decoding Framework That Accelerates DeepSeek-V4 Per-User Generation 60–85% Over MTP-1
DeepSeek搞了个DSpark，让V4推理速度翻倍，开源还无损，适合高并发场景。
DeepSeek开源了DSpark框架，通过将草稿模块附加到现有DeepSeek-V4权重上实现推测解码。它结合并行草稿骨干和轻量级马尔可夫头来减少后缀衰减，并加入基于置信度的调度验证，根据实时GPU负载调整检查token数量。
离线测试中，接受长度相比DFlash和Eagle3提升16-31%；生产环境中每个用户生成速度比MTP-1基线提升57-85%，且无损。
训练代码DeepSpec以MIT许可证开源。
DeepSeek Releases DSpark, a Speculative Decoding Framework That Accelerates DeepSeek-V4 Per-User Generation 60–85% Over MTP-1
DeepSeek open-sourced DSpark, a speculative decoding framework that attaches a draft module to existing DeepSeek-V4 weights. It pairs a parallel draft backbone with a lightweight Markov head to cut suffix decay, then adds confidence-scheduled verification that tailors how many tokens get checked to real-time GPU load. Offline, accepted length rises 16–31% over DFlash and Eagle3; in production it speeds per-user generation 57–85% over the MTP-1 baseline, losslessly. The training repo, DeepSpec, ships under MIT. The post DeepSeek Releases DSpark, a Speculative Decoding Framework That Accelerates DeepSeek-V4 Per-User Generation 60–85% Over MTP-1 appeared first on MarkTechPost .
This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include DSpark, DeepSeek-V4, DeepSeek, 推测解码.
Primary source verification: published directly by marktechpost with first-party authority. Cross-source corroboration: 2 independent sources have covered this development (including IT之家), confirming industry-wide signal strength.
- IT之家06-27 07:47Source
