verl-project/verl

希望可以增加单独检查权重转换逻辑

Open

#6414 opened on May 20, 2026

View on GitHub
 (1 comment) (0 reactions) (0 assignees)Python (21,533 stars) (3,940 forks)auto 404
help wanted

Description

我在尝试把slime的训练任务迁移到verl，但是发现权重转换有点问题，我想对这个进行调试，但是仍然需要足够规模的时间和显存来保证训练和推理的正常进行。请问有什么手段可以单独检查权重转换逻辑嘛？

slime里有--check-weight-update-equal配置来验证权重同步后的值是否正确。https://github.com/THUDM/slime/blob/main/docs/zh/developer_guide/debug.md

Contributor guide

Tech stack
pythonpytorch
Domain
machine learning
Issue type
feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
half day
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
fresh
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
needs investigation
Prerequisites
PythonPyTorchRL training basics
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
25
Research direction
This issue requests a feature to independently verify weight conversion logic, similar to slime's check weight update equal flag. The contributor should first examine the current weight conversion implementation in verl by searching for relevant code (e.g., in files handling model checkpoint loading/saving). They can refer to the slime repository's debug documentation at https://github.com/THUDM/slime/blob/main/docs/zh/developer guide/debug.md for inspiration. After understanding the existing logic, a proposal for a new configuration flag or script to check weight updates can be drafted and discussed in the issue comments with maintainers to clarify the expected behavior and scope.