facebookresearch/fairseq

Use symbolic link for saving best/last checkpoints

Open

#5,490 创建于 2024年4月23日

在 GitHub 查看
 (0 评论) (0 反应) (0 负责人)Python (29,107 star) (6,224 fork)batch import
enhancementhelp wantedneeds triage

描述

🚀 Feature Request

When preparing checkpoint_best.pt and checkpoint_last.pt, create symbolic link instead of making exact copy of the checkpoint

Motivation

I'm running on a machine that somehow has bad io performance, writing the same file 3 times seems extremely inefficient.

Seems that there is plan to implement a asynchronous copying but i believe a symlink is much more efficient https://github.com/facebookresearch/fairseq/blob/bedb259bf34a9fc22073c13a1cee23192fa70ef3/fairseq/checkpoint_utils.py#L128-L141

Pitch

Alternatives

Additional context

贡献者指南

技术栈
pythonpytorch
领域
backendperformance
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。
2
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。
1-3 hours
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。
clear
前置要求
Basic PythonUnderstanding of symbolic linksFamiliarity with fairseq checkpoint logic
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。
80
研究方向
Review the checkpoint saving logic in fairseq/checkpoint utils.py around lines 128-141. Understand how checkpoints are saved and copied. Implement using os.symlink instead of shutil.copy to create symbolic links for best and last checkpoints. Ensure cross platform compatibility (Windows may require additional handling). Test the change locally. Consider edge cases like existing symlinks or permission errors. Discuss any questions in the issue.