KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#dsh #dsh-plugin #ai-agent #developer-tools #codex #typescript #cli #deepseek-harness #claude-code #ai-agents #mcp #rust

infra-skills

★ 140 Open GitHub ↗

Open, portable agent skills for AI infrastructure engineering.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# 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`](skills/accelerators/optimize-musa-training/SKILL.md) | 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`](catalog.json).

## Install

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

```bash
# 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

The category layer keeps a large library navigable while remaining compatible with tools that discover `skills/<category>/<name>/SKILL.md`.

## Scope

Planned categories include:

- `accelerators`: vendor runtimes, device bring-up, profiling, and architecture-aware optimization;
- `distributed-training`: DDP, FSDP, tensor/pipeline parallelism, collectives, and failure diagnosis;
- `inference-serving`: throughput, latency, batching, caching, quantization, and serving engines;
- `compilers-kernels`: graph compilers, custom operators, Triton/CUDA/MUSA kernels, and numerical validation;
- `data-storage`: preprocessing, dataloaders, object storage, checkpoint I/O, and data movement;
- `observability-benchmarking`: MFU/HFU, traces, rooflines, reproducible benchmarks, and cost analysis;
- `cluster-operations`: schedulers, containers, networking, health checks, and incident workflows.

This scope is a direction, not a requirement to create empty scaffolding. Categories are added when a maintained skill exists.

## Design Principles

1. **Evidence first.** Measure before optimizing, and state exactly what a metric includes.
2. **Correctness is a gate.** A speedup is retained only after numerical and behavioral checks pass.
3. **Portable orchestration.** Skills describe target-system behavior without requiring a particular laptop OS, IDE, shell transport, or scheduler.
4. **Progressive disclosure.** Keep the main `SKILL.md` focused and load deeper references only when needed.
5. **Reproducible experiments.** Record source revision, environment, command, workload shape, warmup policy, and raw evidence.
6. **Secure by default.** Never publish credentials, private hosts, proprietary data paths, or internal registry details.
7. **Honest negative results.** Document failed or neutral experiments when they prevent others from repeating expensive work.

## Quality Bar

Every skill must:

- have valid `name` and `description` frontmatter;
- use a lowercase kebab-case directory matching its skill name;
- contain actionable workflows rather than a loose collection of tips;
- separate facts, measurements, inferences, and recommendations;
- include verification steps and failure boundaries;
- avoid private infrastructure details and embedded secrets;
- pass repository validation and keep generated catalog data current.

Run the checks locally with:

```bash
npm test
```

See [`CONTRIBUTING.md`](CONTRIBUTING.md) for the authoring and review workflow.

## Project Status

Open Infra Skills is in early development. The first skill focuses on MUSA training optimization; the repository structure and validation pipeline are designed to support a much broader AI infrastructure collection without turning every skill into one giant prompt.

## License

Licensed under the [Apache License 2.0](LICENSE).

infra-skills

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README