verl-project/verl

[FR] Better metric reduction utility

Open

#1,508 创建于 2025年5月13日

在 GitHub 查看
 (2 评论) (2 反应) (0 负责人)Python (21,533 star) (3,940 fork)auto 404
enhancementfeature requestgood first issue

描述

Motivation

reduce_metrics seems to be used for reducing metrics across DP ranks and batches, which only simply means the values.

However, reducing metrics across DP ranks is non-trivial since some metrics like "mean" are not linear.

贡献者指南

技术栈
pythonpytorch
领域
machine learningdata
议题类型
feature
难度面向新贡献者的预计实现难度，1 表示很小改动，5 表示专家级工作。
4
预计时间有经验贡献者完成调查、实现、测试并准备 pull request 的粗略时间范围。
1-2 days
活动状态议题当前的可参与程度：新鲜、活跃、陈旧、阻塞或等待维护者输入。
fresh
清晰度议题是否清楚说明期望改动、验收标准和下一步。
mostly clear
前置要求
PythonPyTorch distributedverl codebase
新手友好度1-100 的估计分数，表示该议题对首次贡献者的友好程度。
25
研究方向
Investigate the current reduce metrics function at https://github.com/verl project/verl/blob/main/verl/utils/metric/utils.py. Understand the different types of metrics and their reduction requirements across data parallel ranks. Propose a generalized reduction strategy that accounts for non linear metrics like mean. Consider existing implementations in PyTorch or other frameworks.