verl-project/verl

Additional memory optimization features

Open

#144 创建于 2025年1月27日

在 GitHub 查看
 (4 评论) (3 反应) (0 负责人)Python (21,533 star) (3,940 fork)auto 404
call for contributionenhancementgood first issue

描述

  • Activation offloading (see implementation here)
  • Fusing optimizer step into backward pass (see implementation here)
  • Utilize full_shard reshard_after_forward (see here). I wasn't 100% sure if I could see this already implemented in veRL.

These optimizations largely trade off decreased peak memory useage for additional compute, so may only be useful for training larger models, and in GPU-constrained settings.

贡献者指南

技术栈
pythonpytorch
领域
machine learningperformance
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。
over 1 week
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
PythonPyTorchFSDP
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。
15
研究方向
Explore the provided links to activation offloading and fused optimizer step implementations from torchtune (torchtune/training/ activation offloading.py and torchtune/training/memory.py). Study the FSDP2 API differences in torchtitan (docs/fsdp.md). Then, survey the verI codebase to identify where similar memory optimizations can be integrated, focusing on the training loop and model sharding. Consider discussing with maintainers on which optimization is most needed first.