verl-project/verl

Additional memory optimization features

Aberta

#144 aberto em 27 de jan. de 2025

 (6 comentários) (3 reações) (0 responsável)Python (4.440 forks)auto 404
call for contributionenhancementgood first issue

Métricas do repositório

Stars
 (23.087 estrelas)
Métricas de merge de PR
 (Mesclagem média 3d 6h) (125 fundiu PRs em 30d)

Description

  • 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.

Guia do colaborador