Unverified50% confidenceFactExact time
以 GPT-3 为例，r=4 时 LoRA 可将可训练参数从 175B 降至约 4.7M，减少约 37,000 倍
1
Sources
50%
Confidence
Long-term
Relevance
7/14/2026
First Seen
Sources
Related Claims
Unverified一个rank=4的LoRA文件对于SD1.5模型约只有3-6MB，LoRA设计使训练参数量减少95%以上76% similarUnverifiedLoRA使可训练参数量通常仅为全参微调的0.1%-1%，可在单张RTX 3090/4090（显存24GB）上微调7B乃至13B级别的模型72% similarUnverifiedLoRA训练中常用的学习率为1e-4到5e-4之间，常用的Network Rank值为8、16、32、64、12866% similarUnverifiedW4A4量化将权重和激活值均压缩至4位整数，理论上相比FP16可将显存占用降低至约1/463% similarUnverified响应时间只看厂商标称的1ms/0.5ms GtG意义不大，应看RTINGS等第三方实测的过冲(overshoot)数据，标称与实测常差2-3倍62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/503529API
curl https://kongchang.com/api/v1/knowledge/claims/503529MCP
get_claim(id=503529)