Conversation
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request introduces the teacher_model_group argument and refactors the GKDTrainer to use a choose_teacher_model method for selecting teacher models during loss computation. Feedback highlights that the use_mopd flag is missing from the argument definitions, which will result in an AttributeError. Additionally, the choose_teacher_model implementation is currently incomplete and will cause crashes when MOPD is enabled, and the core logic for incorporating teacher weights into the JSD loss calculation is still pending.
| remotely. When this is set, `teacher_model` is not required. Defaults to None. | ||
| """ | ||
| teacher_model: Optional[str] = None | ||
| teacher_model_group: List[str] = field(default_factory=list) |
There was a problem hiding this comment.
use_mopd 标志在 GKDTrainer 中被引用，但未在参数定义中声明。应在此处添加以避免 AttributeError。此外，建议更新 TeacherModelArguments 的 docstring 以包含 teacher_model_group 和 use_mopd 的说明。
| teacher_model_group: List[str] = field(default_factory=list) | |
| teacher_model_group: List[str] = field(default_factory=list) | |
| use_mopd: bool = False |
| def choose_teacher_model(self): | ||
| if not self.args.use_mopd: | ||
| return self.teacher_model | ||
| #todo 使用mopd时从教师模型组选择最佳模型 |
There was a problem hiding this comment.
choose_teacher_model 的实现不完整。当启用 use_mopd 时，该函数目前隐式返回 None，这将导致 compute_loss 在预期有效模型的地方（例如第 311 行或第 419 行）发生崩溃。此外，teacher_model_group 中的模型需要在训练器初始化期间加载并准备为 module 对象，以便在此处使用。
| t_log_probs = F.log_softmax(t_chunk, dim=-1) | ||
| del s_chunk, t_chunk | ||
|
|
||
| #todo 使用mopd的计算函数，增加教师模型权重 |
There was a problem hiding this comment.
此 TODO 表明 MOPD 的核心逻辑（将教师权重纳入 JSD 损失计算）尚未实现。如果没有这部分实现，MOPD 功能将无法按预期工作。
2 participants
PR type
PR information
https://github.com/XiaomiMiMo/MiMo-V2-Flash/blob/main/paper.pdf
小米团队推出的MiMo-V2中提出了一种新的后训练范式Multi-Teacher On-Policy Distillation (MOPD)，用于解决能力不平衡的问题（例如：提升了数学，代码能力下降）。
Write the detail information belongs to this PR.
Experiment results
Paste your experiment result here(if needed).