Event ·Official first-hand

从专家归约到行为分歧:追踪稀疏MoE推理中的数值状态

DeepSeek-V4-Flash的稀疏MoE推理中,数学等价的专家归约顺序会产生可观察的执行分歧。研究对比4种聚合方案:720个A模式顺序产生10个延续盆地,720个B模式顺序形成360个结构类和11个盆地。在一条中文提示词下,B类分裂为202个裁员、113个招聘和45个其他延续。192条持久轨迹中,P32、A、B方案改变所有原生参考路由,而C方案保留路由、token序列和文本。FP64精确重建可在301个下游状态上获得七步一致。

Current read

这篇论文把DeepSeek-V4-Flash的MoE数值细节挖得很深,归约顺序不同结果就不同,跑推理时得留神。

11 sources52° AI heatLast updated 2026/7/30 12:07:56

Evidence

11 sources
Primary sourcearXiv: DeepSeek
View source
Related source 5Fireworks AI
View source
Related source 8深度求索 DeepSeek
View source
Related source 9歸藏(guizang.ai)
View source
Related source 10向阳乔木
View source

Conflict 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.