verl-project/verl

Additional memory optimization features

Open

#144 opened on Jan 27, 2025

View on GitHub
 (4 comments) (3 reactions) (0 assignees)Python (21,533 stars) (3,940 forks)auto 404
call for contributionenhancementgood first issue

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.

Contributor guide

Tech stack
pythonpytorch
Domain
machine learningperformance
Issue type
feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
4
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
over 1 week
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
fresh
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
mostly clear
Prerequisites
PythonPyTorchFSDP
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
15
Research direction
Explore the provided links to activation offloading and fused optimizer step implementations from torchtune (torchtune/training/ activation offloading.py and torchtune/training/memory.py). Study the FSDP2 API differences in torchtitan (docs/fsdp.md). Then, survey the verI codebase to identify where similar memory optimizations can be integrated, focusing on the training loop and model sharding. Consider discussing with maintainers on which optimization is most needed first.