verl-project/verl

[Bug][CI] FDSP2 test in `model_rmpad` job seems unstable

Open

#1,388 创建于 2025年5月4日

在 GitHub 查看
 (0 评论) (1 反应) (0 负责人)Python (21,533 star) (3,940 fork)auto 404
bugcall for contributiongood first issue

描述

Motivation

https://github.com/volcengine/verl/actions/workflows/model.yml shows that:

  1. the FDSP2 test in model_rmpad workflow fails sometimes;
  2. but can also pass sometimes.

Plan

  • Find a setup that can reproduce the error steadily (possibly using the test container)
  • Locate the root cause
  • Fix the bug

Additional Info.

贡献者指南

技术栈
pythondocker
领域
testingdevopsci cd
议题类型
bug
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。
3
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
Pythonunderstanding of CI workflowstest debugging
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。
35
研究方向
First, reproduce the flaky FDSP2 test in the model rmpad job locally using the test container as mentioned in the CI workflow. Inspect logs from successful and failed runs to identify differences. Look at the related PR (#1026) for context. Examine the test code for timing dependencies or race conditions. Check if the failure is environment specific (e.g., GPU or resource contention).