THUDM/slime

[FSDP] Try to optimize context parallelism

Open

#1,062 opened on Dec 8, 2025

View on GitHub
 (6 comments) (0 reactions) (0 assignees)Python (873 forks)github user discovery
good first issue

Repository metrics

Stars
 (6,010 stars)
PR merge metrics
 (PR metrics pending)

Description

Currently the context parallelism's training time is like 2 times when we use cp=2 than cp=1, we should try to optimize this for long sequence, maybe reduce the all gather entropy and logp and calculate the local loss to reduce the numbers of communication

Contributor guide

Research direction
Investigate the FSDP context parallelism implementation to identify communication bottlenecks, specifically the all gather for entropy and logp. Consider computing local loss to reduce communication.
Tech stack
pythonpytorch
Domain
backendmachine learningperformance
Issue type
Performance
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.
Needs investigation
Prerequisites
PythonPyTorch
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
30