Primary sourcearXiv: DeepSeek
View sourceEvent ·Official first-hand
近似投机解码ASD:免训练加速自回归生成
Approximate Speculative Decoding(ASD)是一种免训练的投机解码验证器,用预算化的最长前缀选择替代逐位截断,在保留目标贪心后缀的同时接受部分不匹配token。在Qwen3-14B+DSpark-14B的七项任务上,ASD相比严格验证平均提升7.78%吞吐,区间为3.05%到15.26%。在FP4到FP8兼容设置下,DeepSeek-V4-Flash(284B)搭配DSpark时,GSM8K和MATH-500上的验证方接受率提高约10%到16%。该方法无需新草稿模型或微调,预算为零时精确退化为标准贪心验证,代码已开源。
Current read
这篇论文搞了个叫ASD的投机解码改进,不用重新训练就能提速,实测吞吐最高涨15%,代码也开源了,跑推理的可以看看。
8 sources51° AI heatLast updated 2026/8/4 10:45:24
Evidence
8 sourcesRelated source 1lmarena.ai
View sourceRelated source 2pandaily
View sourceRelated source 3歸藏(guizang.ai)
View sourceRelated source 4IT之家
View sourceRelated source 5Geek
View sourceRelated source 6AI Will
View sourceRelated source 7NVIDIA AI
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.