Verified65% confidenceFactTime unknown
LoRA中低秩矩阵的秩r通常设为8或16，可训练参数量从d×k降至r×(d+k)，降幅可达数百倍
3
Sources
65%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
4-bit QLoRA微调LLaMA 3实战：消费级GPU训练80亿参数大模型指南
githubCre4T3Tiv3
Related Entities
Related Claims
UnverifiedLoRA将可训练参数量从d²降低到2dr，其中r通常为4-6478% similarVerified对于一个4096×4096的权重矩阵，LoRA可能只需要训练两个4096×8的矩阵，将可训练参数量从约1600万降低到约6.5万，减少了99.6%78% similarUnverifiedLoRA将微调权重变化量ΔW分解为两个低秩矩阵的乘积，参数量仅为原模型的0.1%~1%，当r=8、d=4096时参数压缩比高达256倍74% similarUnverified对LLaMA-7B所有注意力层应用r=8的LoRA，可训练参数量从70亿降至约400万，压缩比超过1750倍69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/29159API
curl https://kongchang.com/api/v1/knowledge/claims/29159MCP
get_claim(id=29159)