Unverified50% confidenceSolutionExact time
在8卡H20配置下张量并行度通常设置为8
1
Sources
50%
Confidence
Long-term
Relevance
7/16/2026
First Seen
Sources
Related Claims
UnverifiedLFM2.5-8B-A1B总参数量为8B，推理时仅激活1.5B参数62% similarUnverified在Codex环境下，若用户未做调整，系统默认的并发子线程上限为6，委派层级为157% similarVerifiedMixtral 8x7B模型拥有8个专家，每个Token只激活2个，总参数量达到47B，但单次推理的计算量仅相当于一个13B参数的稠密模型57% similarUnverifiedLaguna S 2.1 总参数为 118B，激活参数仅 8B57% similarUnverified在MoE模型中每个token只激活少量Expert（通常2-4个，Top-K路由），实际计算量可能仅相当于6-7B的Dense模型56% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/533901API
curl https://kongchang.com/api/v1/knowledge/claims/533901MCP
get_claim(id=533901)