alibaba / alibaba/ROLL

Why use DeepSpeed's implementation instead of torch's FSDP?

Open
#20 3 comments 0 reactions 0 assignees View on GitHub
Dominant language
Python
Stars
3.4k
Forks
312
Avg merge
1h 2m
Merged PRs (30d)
2

Description

It's really weird to see that a training framework in 2025 is still using DeepSpeed.
DeepSpeed's performance is terrible, and they nearly give up maintaining the repo themselves.
I can't see why you guys are still using DeepSpeed as the backend for ZeRO.

Contributor guide

No contributing guide indexed for this repository

Research direction

Start by locating the ZeRO backend and comparing its DeepSpeed integration with torch's FSDP. A complete outcome would document the performance and maintenance rationale and, if a change is agreed, define the migration scope and validation criteria.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
distributed-systems, machine-learning
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
20/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.