Skip to content

README.md

Latest commit

 

History

History
193 lines (130 loc) · 13.7 KB

README.md

File metadata and controls

193 lines (130 loc) · 13.7 KB
Raw

AMD Skills

AMD Skills give coding agents the knowledge, scripts, and conventions they need to work with AMD hardware and software. Each skill follows the standardized Agent Skills format and works with Cursor, Claude Code, OpenAI Codex, and Gemini CLI.

Installation

AMD Skills is built directly into Claude and Cursor. No install. No setup.

Just ask something like: "Use AMD Skills to integrate local AI into my app".

For other agents, see Manual installation.

What is a skill?

A skill is a self-contained folder that bundles everything an agent needs to perform a focused task: instructions, helper scripts, prompts, templates, and references. At its core is a SKILL.md file with YAML frontmatter, a name, and a short description that tells the agent when the skill should activate, followed by the guidance the agent reads while the skill is in use.

skills/
  rocm-doctor/
    SKILL.md
    scripts/
    references/

When an agent decides a skill is relevant (or you invoke it explicitly), it loads that SKILL.md and follows the instructions inside. Descriptions stay in context cheaply; the full body of a skill only loads when the task actually matches.

Why a skill, not a doc?

Documentation describes an API surface: every flag, every option, neutral by design. A skill encodes the opinionated path: which flags, which container image, which gfx target, which environment variables, in what order. It captures the decisions a senior AMD engineer makes without thinking, in a form the agent can apply consistently across teams and repositories.

Skills earn their keep on repeated, opinionated workflows, exactly where the AMD stack lives.

The catalog

The initial catalog is organized into five focus areas.

Application integration

Embed AMD-optimized AI into end-user applications.

Skill What it does Source
local-ai-app-integration Integrate local AI into cloud LLM apps for offline support, better privacy, and lower API costs. in-repo
local-ai-use Route image generation, text-to-speech, and speech-to-text through a local AI server to reduce token cost. in-repo

Hardware-native skills

Diagnose, configure, and tune AMD devices directly.

Skill What it does Source
apu-memory-tuner Inspect and tune the shared-vs-dedicated memory split (GTT / UMA Frame Buffer) on AMD Ryzen APUs. in-repo
rocm-doctor Diagnose ROCm / PyTorch / llama.cpp failures on AMD GPUs against a fixed list of known misconfigurations. in-repo
mi-tuner Opinionated inference tuning for MI accelerators (TunableOp, FSDP, FlashAttention). planned
gfx-target-chooser Pick the right gfx942 / gfx90a / gfx1100 target and matching compiler flags. planned

Kernel optimization

Write, tune, and reason about GPU kernels for AMD targets. All entries are federated from AMD-AGI/Apex at main (tools/skills/).

Skill What it does Source
aiter-reflection Optimize AMD GPU kernels on MI300 using the aiter project: op tests, benchmarks, iteration, experiment database. Apex
gpu-architecture-fundamentals Reason about memory hierarchy, execution model, block sizing, and latency across HIP, Triton, and PyTorch. Apex
hip-kernel-optimization Write and tune HIP kernels: coalescing, shared-memory tiling, bank conflicts, warp primitives, occupancy, vectorization. Apex
kernel-exp-history Consult past kernel optimization experiments and record the current iteration back into the experiment database. Apex
mi300-hip-programming-insights CDNA3 / MI300 HIP programming insights: chiplet and cache model, Infinity Cache, coherency, matrix cores, sparsity. Apex
pytorch-kernel-optimization Optimize PyTorch models and kernels: torch.compile, custom extensions, mixed precision, CUDA graphs, profiling. Apex
triton-hip-reference-kernel-search Search and adapt Triton / HIP kernel patterns from a corpus to reuse tiling and occupancy strategies. Apex
triton-kernel-optimization Write and tune Triton kernels: autotune block sizes, tiled matmul, fused ops, reductions, flash-attention, quantization. Apex
triton-kernel-reflection-prompts Reflection / self-critique prompts for reviewing and fixing AMD-targeted Triton kernels. Apex

Cross-stack porting

Bring existing workloads onto AMD.

Skill What it does Source
cuda-to-hip Port CUDA kernels with hipify and flag anything that needs manual review. planned
vllm-rocm Stand up vLLM on AMD with the right environment variables and model configurations. planned
pytorch-rocm-setup Get a known-good PyTorch + ROCm stack running on a target node, end to end. planned

Profiling and delivery

Close the loop from trace to fix to ship.

Skill What it does Source
rocprof-compute Profile AMD GPU kernels with rocprof-compute to collect metrics, roofline data, and bottleneck analysis. Apex
rocprof-capture Capture and interpret a rocprof trace for a workload. planned
omniperf-tune Run omniperf, locate the bottleneck, and suggest the fix. planned
quark-quantize Quantize PyTorch / ONNX models with AMD Quark and export for AMD deployment. planned

A federated catalog

The AMD stack is large and moves fast. ROCm, HIP, Ryzen AI, and framework integrations each have their own team, release cadence, and validation matrix. So skills here are federated: each skill is owned and versioned by the team that owns the product it describes, and this repository is the catalog that brings them together.

                ┌─────────────────────────────────────────────────────┐
                │                amd/skills (this repo)               │
                │                                                     │
                │   skills/         scripts/         .*-plugin/       │
                │   in-repo skills  sources.yml      agent manifests  │
                └──────────────────────┬──────────────────────────────┘
                                       │  one install
                                       ▼
                              your AI coding agent
                                       ▲
                                       │  resolves pointers to
       ┌───────────────┬───────────────┼───────────────┬────────────────┐
       │               │               │               │                │
   ROCm/ROCm       ROCm/HIP        Ryzen AI repo   lemonade-sdk    ...more
  rocm-doctor/    cuda-to-hip/    ryzen-ai-tools/   local-ai-app-   product
   gfx-target-...  triton-amd-...  ...               integration/    repos

This repo also acts as an incubator: a skill can start under skills/ to iterate quickly, then graduate to its product repo and be re-pointed from scripts/sources.yml once it has a clear owner, with no change for installed users.

  • One install, full coverage. Add this repository through your agent's plugin flow and you get the whole AMD catalog.
  • Skills update with the products they describe. When ROCm cuts a release, the ROCm team updates the ROCm skills as part of that release.
  • Skills you can trust. Each skill is signed off by the team that owns the underlying product.

Each vendored skill carries a .federated.json marker that records the upstream repo and pinned commit, so the importer can refresh or remove it without disturbing in-repo skills.

Manual installation

AMD Skills are compatible with Cursor, Claude Code, OpenAI Codex, and Gemini CLI.

Cursor

Install the AMD plugin from this repository through the Cursor plugin flow. The repo ships a .cursor-plugin/plugin.json so skills are discoverable as soon as the plugin is enabled.

Claude Code

Register this repository as a plugin marketplace, then install individual skills:

/plugin marketplace add amd/skills
/plugin install <skill-name>@amd/skills

OpenAI Codex

Copy or symlink the desired folders from skills/ into one of Codex's standard skill locations (for example $REPO_ROOT/.agents/skills or $HOME/.agents/skills). Codex discovers SKILL.md files automatically.

Gemini CLI

A gemini-extension.json will be provided so the repo can be installed as a Gemini CLI extension:

gemini extensions install https://github.com/amd/skills.git --consent

Using a skill

Reference it in plain language while talking to your agent. The agent loads the matching SKILL.md and any helper scripts, then carries out the task.

  • "Use AMD Skills to integrate local AI capabilities into my app with Embeddable Lemonade."
  • "Use AMD Skills to convert these CUDA kernels and flag anything that needs manual review."

In most cases the agent picks the right skill on its own from the description; explicit invocation is a fallback, not a requirement.

Contributing a skill

We welcome contributions from AMD engineers and selected partners. Two paths, matching how the catalog is organized:

  • Path A — In-repo skills. Authored directly under skills/. Best for cross-cutting workflows without a natural product home.
  • Path B — Product-repo skills. Authored in a product repository and registered here through scripts/sources.yml with a pinned tag. Best for skills that should ship and version with a specific product.

See CONTRIBUTING.md for step-by-step instructions and the rules CI enforces.

License

Released under the MIT License. See LICENSE for details.