Primary sourceelvis
View sourceEvent ·Multi-source
NVIDIA 推出 LoGRA:显存占用低于 LoRA 的后训练 RL 方法
NVIDIA 发布 LoGRA 论文,一种面向 LLM 后训练 RL 的省显存技术。在 1.5B 模型上,LoGRA 平均训练显存 7.2 GiB,低于 rank-256 LoRA 的 13.2 GiB,Pass @4 两者接近(81.8% 对 81.4%)。其原理是用每步重绘的 rank-256 随机投影压缩权重梯度,不构建完整梯度,并用 KL 预算约束策略更新幅度。在 7B 模型上,每 GPU 平均训练显存下降 45.7%(31.8 降至 17.3 GiB),Pass @1 精度 72.3% 接近 dense Adam 的 72.5%。该方法还可在单个八 GPU 节点上对 27B 模型训练超过 1,100 步,而 dense Adam 在此配置下会显存不足。
Current read
显存不够跑 RL 的朋友看过来,NVIDIA 这个 LoGRA 比用 LoRA 还省显存,7B 模型省了近一半,精度基本不掉。
11 sources38° AI heatLast updated 2026/10/12 22:20:02
Evidence
11 sourcesRelated source 1IT之家
View sourceRelated source 2The Straits Times: Business
View sourceRelated source 3Yuchen Jin
View sourceRelated source 4arXiv cs.LG
View sourceRelated source 5rohanpaul_ai
View sourceRelated source 6VTV Công nghệ
View sourceRelated source 7vLLM
View sourceRelated source 8DavidSacks
View sourceRelated source 9NVIDIA
View sourceRelated source 10Clement Delangue
View sourceConflict 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.