THUDM/slime

[FSDP] Try to optimize context parallelism

Open

#1,062 创建于 2025年12月8日

在 GitHub 查看
 (6 评论) (0 反应) (0 负责人)Python (873 fork)github user discovery
good first issue

仓库指标

Star
 (6,010 star)
PR 合并指标
 (平均合并 1天 11小时) (30 天内合并 31 个 PR)

描述

Currently the context parallelism's training time is like 2 times when we use cp=2 than cp=1, we should try to optimize this for long sequence, maybe reduce the all gather entropy and logp and calculate the local loss to reduce the numbers of communication

贡献者指南