Skip to content

support weight-update in disaggregated mode using sglang#1766

Open
PengchengShi00 wants to merge 1 commit intoInternLM:rl_designfrom
PengchengShi00:support-weight-update-sglang-dis
Open

support weight-update in disaggregated mode using sglang#1766
PengchengShi00 wants to merge 1 commit intoInternLM:rl_designfrom
PengchengShi00:support-weight-update-sglang-dis

Conversation

@PengchengShi00
Copy link
Copy Markdown

  1. 将TrainingWorker中有关权重同步的函数抽取到UpdateWeighter类中
  2. 将PR中有关训推共卡权重同步的优化更新到UpdateWeighter类中,Update weight persist buffer #1653
  3. 修复sglang在跑训推分离时,rollout和train所占GPU没有分开部署的 bug
  4. 更新配置GSM8KJudgerConfig的配置参数
  5. 增加训推分离模式下,使用sglang作为推理后端时的权重同步
    a. 创建训练 ranks 之间使用的 gloo group,训推分离权重同步时通过该group做 barrier
    b. 创建了一个 NCCL process group,用来将训练 rank0 把 bucket 后的权重 broadcast 给 SGLang rollout ranks:

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant