amd/skills

What is AMD Skills?

AMD Skills is a catalog of “skills” – small, self‑contained bundles of instructions, scripts, prompts and metadata – that let AI coding assistants (Cursor, Claude Code, OpenAI Codex, Gemini CLI, etc.) work fluently with AMD’s hardware and software stack. Each skill tells an agent when it should be used (e.g., “deploy an LLM on an Instinct GPU”) and then provides the exact commands, environment variables, and best‑practice steps an AMD engineer would follow.

The repository does not contain the AI models themselves; it only ships the knowledge needed to install, configure, diagnose and optimise AMD‑centric AI workloads such as ROCm, Ryzen AI, AMD Instinct, and the AMD‑Quark quantisation tool.

---

How to get the catalog

# Install via the generic “skills” CLI (requires Node.js)
npx skills add amd/skills

The CLI will prompt you to pick one or more individual skills and the target agent (e.g. --agent claude-code). You can also list everything first with --list or install a single skill directly:

npx skills add amd/skills --skill rocm-doctor --agent cursor

If you prefer a manual approach, clone the repo and copy the desired skills/<name>/ folder into the agent’s personal or project‑level skills directory (paths listed in the README).

---

Using a skill

Once installed, you simply talk to your coding agent in natural language, for example:

• “Use AMD Skills to generate images locally instead of calling the cloud API.”
• “Deploy this LLM on my AMD Instinct GPU using the AMD Skills catalog.”

The agent will automatically match the request to the appropriate skill based on the SKILL.md description. If the match fails you can force it with an explicit invocation, e.g. Use AMD Skills: rocm-doctor.

---

The catalog at a glance

Area	Example skill	What it does
Client‑native	local-ai-use	Routes image generation, TTS, STT through a local AMD‑powered AI server to cut cloud token costs.
	local-ai-app-integration	Adds offline AI capability to cloud‑based LLM apps for privacy and cost savings.
Cross‑stack	rocm-doctor	Diagnoses and fixes common ROCm/HIP/PyTorch/llama.cpp mis‑configurations on AMD GPUs.
	lemonade-router-builder	Builds a content‑aware model router (Lemonade) that directs requests based on sensitivity or required capabilities.
	quark-install	Installs or verifies AMD Quark‑accelerated PyTorch builds from PyPI, wheels or source and validates the runtime.
	quark-torch-llm-ptq	Performs post‑training quantisation (FP8, INT4, …) of Hugging‑Face / PyTorch LLMs with AMD Quark and produces a verified model.
	hrr-replay-analysis (planned)	Records, replays and analyses GPU workload traces across Instinct, Radeon and Ryzen hardware.
Server‑native	serving-llms-on-instinct	End‑to‑end deployment of LLM inference on AMD Instinct GPUs, including hardware detection, vLLM recipe selection and benchmarked endpoint launch.
	serving-llms-on-epyc	Serves LLMs on AMD EPYC CPUs with vLLM/zentorch inside Docker/Podman or Conda, handling CPU detection, RAM checks and health monitoring.
	hyperloom-workload-optimizer	Sets up Hyperloom to automatically optimise LLM throughput on Instinct GPUs and reports the gain.
	magpie-kernel-evaluator	Benchmarks and validates GPU kernels, comparing implementations and profiling vLLM/SGLang runs.
	tracelens-analysis-orchestrator	Orchestrates TraceLens‑based PyTorch profiling, generating performance reports and stakeholder summaries.

---

What a skill looks like

Each skill lives in skills/<skill‑name>/ and contains at minimum:

SKILL.md          # YAML front‑matter + description used by agents
skill-card.md    # Human‑readable governance info (owner, license, etc.)
scripts/          # Optional helper scripts
references/      # Optional docs, links, datasets

The SKILL.md tells the agent when the skill applies and provides the step‑by‑step guidance the agent will execute.

---

Governance & contribution

• Federated catalog – Skills are sourced from the owning product repos (ROCm, Quark, Lemonade, etc.). A nightly federate-skills workflow copies only changed skill folders, updates manifests and opens a PR for each bump, ensuring every change is reviewed.
• Contribution flow – AMD engineers or approved partners add new skills by forking, editing the skill folder, and opening a PR. CI validates the skill against the amd/skillscope grader and checks compliance with the docs in docs/.
• License – MIT, with a copyright notice for AMD (2026).

---

Who might use this?

• AI developers who want their coding assistants to automatically handle AMD‑specific setup (e.g., ROCm driver issues, GPU quantisation, LLM serving) without memorising low‑level commands.
• Data‑science teams deploying models on AMD Instinct or EPYC servers who need repeatable, opinionated deployment scripts.
• Tool‑makers building custom agents that can tap into a shared, version‑controlled knowledge base for AMD hardware.

---

Bottom line

AMD Skills turns expert AMD‑hardware know‑how into machine‑readable packages that AI agents can invoke on demand, dramatically lowering the friction of developing, debugging and scaling AI workloads on AMD GPUs, CPUs and Ryzen AI accelerators.