MCP App StoreMCP App Store
from megatron-bridge637
Guides enabling and validating MoE expert-parallel communication overlap in Megatron-Bridge to hide dispatch/combine latency and improve throughput.
This skill explains how to enable and validate Expert-Parallel (EP) overlap for Mixture-of-Experts (MoE) models in Megatron-Bridge. It describes configuration options (overlap_moe_expert_parallel_comm, delay_wgrad_compute), dispatcher choices (alltoall vs flex), minimal working configs, runtime constraints, and verification steps including unit tests and log checks. The guidance is practical: start with the alltoall dispatcher for compatibility, then move to flex (DeepEP/HybridEP) on supported GPUs for higher overlap.
Use this skill when running MoE models where dispatch/combine communication is a measurable bottleneck and you want to trade memory/time to improve throughput. It's appropriate for tuning throughput on multi-GPU nodes, diagnosing throughput regressions related to EP settings, or validating complex interactions (delayed wgrad, CUDA graphs). Avoid using it during early bring-up, or when full activation recompute is enabled.
alltoall and flex dispatchers, minimal runnable commands, verification steps, and known failure modes and fixes.This is a technical, code-focused skill useful to agents that can read code, run tests, and edit training configs (Copilot-style assistants, developer-facing agents). It assumes access to a developer environment with Python and the training repo.
NVIDIA Megatron-Bridge MoE expert-parallel overlap skill. The SKILL.md content is no longer accessible (404 on GitHub), and no scripts were bundled. Only metadata (name, description, tags) was available for review. Very niche — targets Megatron-LM MoE training performance tuning, relevant to only a small subset of ML engineers working with NVIDIA's distributed training framework.
SKILL.md body was null in DB and GitHub returns 404. Skill may be deleted from source repo. Architecture score low due to missing content and no structure to evaluate. Quality score low because there's nothing substantive to review — only a short_description and tags. Usefulness is niche (MoE overlap tuning for Megatron). No security concerns found in available metadata.
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.