verl-project/verl

[FR] Tests for gradient accumulation

Open

#907 建立於 2025年4月3日

在 GitHub 查看
 (0 留言) (0 反應) (0 負責人)Python (21,533 star) (3,940 fork)auto 404
enhancementgood first issue

描述

Summary

Promise from #906.

Add tests to assert the losses on the same mini-batch are equal

  • between w/o & w/ gradient accumulation
  • using all the different loss aggregation modes

Plan

We might went to test end-to-end to avoid potential implication by any other changes.

Plan Pro
A) Inherit a new Trainer class to test Standalone without affecting the main config
B) Add a config term and modify update_* to test 1) Minimal code modification 2) Compatibility to any Trainer

貢獻者指南

技術棧
pythonpytorch
領域
testing
議題類型
feature
難度面向新貢獻者的預計實作難度，1 表示很小改動，5 表示專家級工作。
3
預計時間有經驗貢獻者完成調查、實作、測試並準備 pull request 的粗略時間範圍。
half day
活動狀態議題目前的可參與程度：新鮮、活躍、陳舊、阻塞或等待維護者輸入。
fresh
清晰度議題是否清楚說明預期改動、驗收標準和下一步。
clear
前置要求
PythonPyTorchgradient accumulation
新手友善度1-100 的估計分數，表示該議題對首次貢獻者的友善程度。
70
研究方向
Investigate the two proposed plans: A) creating a new Trainer subclass for testing to avoid affecting main config, or B) adding a config term and modifying update * methods. The tests should assert that losses on the same mini batch are equal with and without gradient accumulation across all loss aggregation modes. Look at the existing test suite in the repository to understand the testing structure and consider implementing both options if feasible. Examine the gradient accumulation implementation in the trainer to identify the update * methods and loss aggregation modes.