Event ·Multi-source

NVIDIA NeMo Megatron Bridge 支持离线参考打分的 DPO 训练

NVIDIA 在其训练库 NeMo Megatron Bridge 中实现了 DPO 训练,与 SFT 一样只用一份模型,显存占用和步进耗时不变。传统 DPO 需要在每个训练步额外运行一个冻结的参考模型,该实现改为离线打分一次,把每对 chosen/rejected 的 logprob 总和存到磁盘,训练时直接读取。文件保存的是未加 beta 的原始总和,beta 和学习率在损失计算时才引入,一次打分即可覆盖后续所有超参数搜索。

Current read

DPO 训练不用再挂两份模型了,NVIDIA 把参考打分挪到离线做一次,显存和速度跟 SFT 一个样,调参还不用重算。

10 sources43° AI heatLast updated 2026/11/4 12:53:00

Evidence

10 sources
Related source 1TestingCatalog
View source
Related source 2kimmonismus
View source
Related source 4Fireworks AI
View source
Related source 7OpenRouter
View source

Conflict 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.

NVIDIA adds 1/2 DPO training to NeMo Megatron Bridge · 事件专题 | AITOP