为什么值得关注

它将可重复的、基于证据的工程工作流封装为独立的智能体技能，强调可复现性、正确性验证以及跨多种 AI 基础设施环境的可移植编排。

适合谁使用

AI 基础设施工程师
机器学习性能工程师
AI 系统集群运维人员
从事 GPU/加速器部署与优化的开发者

典型使用场景

在系统变更前建立可复现的基线
对 MUSA GPU 上的分布式训练进行性能剖析和优化
跨应用、系统追踪和内核层级诊断性能问题
在提升性能的同时验证数值正确性

项目优势

基于证据的设计，明确测量并区分事实与假设
支持本地、SSH、调度器、容器和云环境，不依赖客户端操作系统
渐进式信息展示：主技能文件简洁，仅在需要时加载深层参考资料
质量标准要求可操作的工作流、验证步骤以及诚实的负面结果记录

使用前须知

处于早期开发阶段，目前仅有一个已发布的技能（MUSA 训练优化）
依赖 Skills CLI 及兼容的智能体，限制了在 Open Agent Skills 生态系统之外的直接采用
当前加速器范围较窄；其他类别（分布式训练、推理等）尚在规划中，尚未实现

README 快速开始

Open Infra Skills

Open, portable agent skills for building, operating, profiling, and optimizing AI infrastructure.

This repository is intended to grow into a comprehensive, vendor-aware library for the full AI systems lifecycle: accelerator bring-up, distributed training, inference serving, kernels and compilers, data pipelines, observability, benchmarking, and cluster operations.

Why This Repository Exists

AI infrastructure work is full of environment-specific details, expensive experiments, and performance claims that are easy to mismeasure. Open Infra Skills packages repeatable engineering workflows as self-contained agent skills so that an assistant can:

establish a reproducible baseline before changing a system;
choose the right diagnostic layer, from application timing to system traces and kernel counters;
preserve numerical behavior while optimizing performance;
distinguish measured evidence from hypotheses;
work across local machines, SSH hosts, schedulers, containers, and clouds without assuming one client operating system.

Skill Catalog

Skill	Category	What it helps with
`optimize-musa-training`	Accelerators	Profile, debug, and optimize AI training on Moore Threads MUSA GPUs, including MFU/HFU, Moore Perf, FA2, `torch.compile`, FSDP, MCCL, and kernel-level analysis.

The machine-readable catalog is available in catalog.json.

Install

The repository follows the open Agent Skills directory convention. Use the Skills CLI to inspect or install it:

# List all skills in the repository.
npx skills add open-infra-skills/infra-skills --list

# Install one skill interactively.
npx skills add open-infra-skills/infra-skills --skill optimize-musa-training

# Install one skill globally for Codex without prompts.
npx skills add open-infra-skills/infra-skills \
  --skill optimize-musa-training \
  --global \
  --agent codex \
  --yes

You can also clone the repository and point a compatible agent at a skill directory directly. Each skill keeps its instructions and optional scripts, references, and agent metadata together.

Repository Layout

infra-skills/
|-- skills/
|   `-- /
|       `-- /
|           |-- SKILL.md
|           |-- agents/       # Optional agent-facing metadata
|           |-- references/

项目描述

Open, portable agent skills for AI infrastructure engineering.

infra-skills