THUDM/slime

[FSDP] Try to optimize context parallelism

Open

#1,062 opened on 2025年12月8日

GitHub で見る
 (6 comments) (0 reactions) (0 assignees)Python (873 forks)github user discovery
good first issue

Repository metrics

Stars
 (6,010 stars)
PR merge metrics
 (平均マージ 1d 11h) (30d で 31 merged PRs)

説明

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

コントリビューターガイド