Unverified80% confidenceFactTime unknown

SVDQuant中低秩分支的参数量通常只占总参数的几个百分点

1

Sources

80%

Confidence

Long-term

Relevance

6/1/2026

First Seen

Sources

SVDQuant: 4-bit Quantization Enables Efficient Diffusion Model Inference on Consumer GPUs

githubnunchaku-ai

Related Entities

SVDQuantconcept

Related Claims

Unverified单条推理通常只能占用GPU SM利用率的5%～15%，而合理的批次大小可将利用率提升至60%～90%64% similar Unverified轻量版模型单次推理的算力消耗往往只有旗舰版本的十分之一甚至更低63% similar Unverified马尔可夫头仅需记录前一个token状态即可推导下一token修正概率，参数量通常不超过原模型的0.1%62% similar UnverifiedPER中重要性采样权重的β参数从较小值（如0.4）线性退火到161% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/28970

API

curl https://kongchang.com/api/v1/knowledge/claims/28970

MCP

get_claim(id=28970)