MoE Expert-Parallel Overlap (Megatron-Bridge)

Name: MoE Expert-Parallel Overlap (Megatron-Bridge)
Rating: 59 (1 reviews)
Author: nvidia

Trust Score 59/100

from skills1,118

Guidance and configs to enable expert-parallel communication overlap in Megatron-Bridge for MoE models — use to hide dispatch/combine latency and improve throug

triggers:moeexpert-paralleloverlapdelay_wgrad_computedeepEPhybridepflex dispatcheralltoall

GitHub SKILL.md

What it does

Provides detailed guidance, example configs, and verification steps for enabling expert-parallel (EP) communication overlap in Megatron-Bridge. Describes dispatcher choices (alltoall vs flex), delayed weight-gradient computation, backend constraints (DeepEP/HybridEP), and minimal working configs to safely roll out overlap.

When to use it

Use when running MoE models where expert dispatch/combine all-to-all communication is a measurable bottleneck and you have the memory and deployment constraints to tune for throughput. Avoid for tiny runs, early correctness bring-up, or incompatible PyTorch/TE/CUDA setups.

What's included

Scripts: none in this skill package, but it references performance harness commands and unit tests in the repo.
References: structured docs and a card.yaml with metadata.
Instructions: config snippets, compatibility matrix, verification steps (unit tests, log checks), minimal runnable commands and failure diagnosis.

Compatible agents

Engineers and agents with knowledge of deep-learning training infra (Megatron/Bridge) — useful for performance-tuning assistants and infra automation tools.

Audit Summary

NVIDIA Megatron-Bridge expert-parallel overlap tuning skill. No scripts bundled. The SKILL.md body is null and the GitHub source URL returns 404, making the actual content inaccessible. Appears to be a broken or removed skill entry with only metadata (slug, tags, truncated description) available. Cannot verify any guidance content.

Watch Out

SKILL.md body is null — content unavailable
GitHub source URL returns 404
Description is truncated in DB

Notes

Skill content appears to have been removed or the URL is incorrect. The DB entry has metadata but no actual skill body. Recommend re-crawling or marking as broken.

Information

Repository: skills
Stars: 1,118
Installs: 0

Trust Score

Overall59

Security95

Code Quality25

Architecture20

Usefulness38

More from skills

Megatron CI/CD Guide

Reference guide to Megatron-LM CI/CD: pipeline structure, PR scope labels, triggering internal CI, and steps for investigating CI failures.

Video Analytics (VA-MCP)

Query incidents, alerts, sensor counts and metrics from a VA-MCP Elasticsearch backend (port 9901) to answer questions about violations, occupancy, speeds, and

VSS Video Summarization

Summarize recorded video clips using a local LVS summarization microservice with HITL; fallback to a VLM when the service is unavailable.

Related Skills

OpenTestAI

Automated, high-confidence AI testing: bug detection, persona feedback, and prioritized test-case generation using many specialized tester profiles.

Go Data Structures

Authoritative guidance on choosing and using Go built-in and standard-library data structures, with practical best practices for slices, maps, arrays, container

React Development Expert

Provides authoritative React guidance on hooks, state patterns, Server Components, performance optimization, and common architectural patterns.

Code Reviewer

Perform structured, prioritized code reviews that find correctness, security, performance, reliability, and testing issues and provide concrete fix suggestions.

dotLottie Web

Guidelines and patterns for implementing performant dotLottie/Lottie animations on the web (vanilla JS and React), including workers, state machines, and themin

Synalinks Framework

Keras-inspired framework for building structured, neuro-symbolic LLM programs with DataModel schemas, modular Programs, and training/optimization tools.

Party Engine Skill

Guidance and examples for using the @cazala/party particle engine (engine lifecycle, modules, WebGPU vs CPU patterns) in custom apps.

Party Skill

Programmatic guide for the @cazala/party library: engine setup, modules, particle APIs, and performance tips for WebGPU and CPU runtimes.

Back to Skills