debug-distributed-hang

Official

Find and fix distributed hangs in SGLang.

AuthorOpenMOSS
Version1.0.0
Installs0

System Documentation

What problem does it solve?

Debug hanging issues in multi-GPU SGLang distributed inference by providing a structured workflow to locate divergence across ranks during collectives and restore progress.

Core Features & Use Cases

  • Per-rank logging patterns to capture divergence and enable diffing across ranks
  • Guided use of watchdogs, NCCL logs, and CUDA coredumps to identify hang causes
  • Binary-search methodology to locate the first point of divergence and apply fixes in TP/PP/DP/EP configurations
  • Use cases include diagnosing AllGather/AllReduce deadlocks and mismatched inputs in model-parallel inference

Quick Start

Run a multi-GPU SGLang run with per-rank logging enabled and follow the steps to locate the first diverging point and implement the resolution.

Dependency Matrix

Required Modules

None required

Components

Standard package

💻 Claude Code Installation

Recommended: Let Claude install automatically. Simply copy and paste the text below to Claude Code.

Please help me install this Skill:
Name: debug-distributed-hang
Download link: https://github.com/OpenMOSS/MOSS-VL/archive/main.zip#debug-distributed-hang

Please download this .zip file, extract it, and install it in the .claude/skills/ directory.
View Source Repository

Agent Skills Search Helper

Install a tiny helper to your Agent, search and equip skill from 471,000+ vetted skills library on demand.