verl-project/verl

Additional memory optimization features

開放

#144 建立於 2025年1月27日

 (6 則留言) (3 個反應) (0 位負責人)Python (4,440 個分叉)auto 404
call for contributionenhancementgood first issue

倉庫指標

星標
 (23,087 顆星)
PR 合併指標
 (平均合併 3天 6小時) (30 天內合併 125 個 PR)

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.

貢獻者指南