SFT 和 DPO 两阶段Lora训练，如何训练同一个 adapter

Checklist / 检查清单

• I have searched existing issues, and this is a new question or discussion topic. / 我已经搜索过现有的 issues，确认这是一个新的问题与讨论。

Question Description / 问题描述

swift rlhf
--rlhf_type dpo
--train_type lora
--model /cpfs_fundata/baolujia.blj/models/Qwen3-8B
--resume_from_checkpoint v0-20260212-195919/checkpoint-4578
--resume_only_model
--ignore_data_skip \

和 直接在合并之后的模型上 lora
swift rlhf
--rlhf_type dpo
--train_type lora
--model v0-20260212-195919/checkpoint-4578-merged
--resume_only_model
--ignore_data_skip \

loss完全不一样

Start by reproducing the loss difference between resuming the LoRA checkpoint with --resume_from_checkpoint and training LoRA on the merged checkpoint using the commands in the issue. Trace how SFT/DPO checkpoint resumption and merged-model loading are handled, and consider the investigation complete when the differing behavior is explained or a reproducible fix is identified.