verl-project/verl

[FR] Documentation about Resolving OOM

Closed

#1,014 opened on Apr 10, 2025

View on GitHubContributor guide
 (6 comments) (7 reactions) (0 assignees)Python (4,413 forks)auto 404
call for contributiondocumentationgood first issue

Repository metrics

Stars
 (22,985 stars)
PR merge metrics
 (Avg merge 3d 6h) (125 merged PRs in 30d)

Description

Motivation

There are many issues related to OOM, e.g. #328 . We might need a clear guide about how to resolve OOM.

Plan

A non-exclusive enumeration about related configurations:

  1. Rollout：gpu_memory_utilization
  2. Other Inference：
    1. Liger Kernel
    2. *_max_len_per_gpu / micro_batch_size_per_gpu
  3. Training:
    1. Liger Kernel
    2. Ulysses Sequence Parallelism
    3. gradient checkpointing
    4. offload

TODO

  • Complete the list of related configurations
  • Benchmark the effect & overhead of each configuration

Contributor guide

Research direction
Investigate existing OOM related issues, e.g., #328, to identify common scenarios. For each configuration listed (rollout's gpu memory utilization, Liger Kernel, max len/gpu, gradient checkpointing, etc.), benchmark their effect on memory usage and training performance. Document findings in a structured guide, referencing relevant source files like configuration scripts and training loops.
Tech stack
pythonpytorch
Domain
documentationperformance
Issue type
Documentation
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
1-2 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Clear
Prerequisites
basic Pythonfamiliarity with PyTorchunderstanding of GPU memory
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
70