THUDM / THUDM/slime

CPU offloading in slime is too aggressive

Open
#113 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
8.5k
Forks
1.3k
Avg merge
5h 36m
Merged PRs (30d)
22

Description

It seems that CPU offloading in SGLang utilizes torch_memory_saver and that in Megatron/slime uses cumem_allocator. There is only a global switch that turns on offloading or not, which wastes a massive amount of host memory when GPU memory is sufficient, especially when training small models (e.g. 0.5B-7B). Fine-grained offloading may lower the host memory requirement for training RL models.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by tracing the global offloading switch in the Megatron/slime path and comparing its use of cumem_allocator with SGLang's torch_memory_saver. Define which model components should be offloaded selectively, then measure host-memory use on the stated 0.5B–7B training cases while preserving RL training behavior.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
machine-learning
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.