Skills catalog

This page lists the skills currently available in the AMD Skills repository, organized by focus area.

The initial catalog is organized into three focus areas, spanning the full stack from client to cloud. This catalog is expected to grow significantly as more skills land.

Client-native

Run and optimize on Ryzen AI.

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

Cross-stack

Cross-stack skills, from client to cloud.

Skill	What it does	Source
rocm-doctor	Diagnose and fix common ROCm, HIP, PyTorch, and llama.cpp failures and misconfigurations on AMD GPUs.	rocm-cli
lemonade-router-builder	Set up a Lemonade model router that handles requests based on content, sensitivity, or required capabilities.	Lemonade
quark-install	Install or verify AMD Quark with an accelerator-matched PyTorch, from PyPI, a native-wheel index, a local wheel, or source, then verify imports, kernels, and quark-cli.	Quark
quark-torch-llm-ptq	Quantize PyTorch / Hugging Face LLMs with AMD Quark post-training quantization: inspect the model, pick a scheme (FP8, INT4, and more), and produce a verified quantized model.	Quark

Server-native

Run and optimize on AMD Instinct.

Skill	What it does	Source
serving-llms-on-instinct	Deploy LLM inference on AMD Instinct GPUs end-to-end: detect hardware (or onboard via AMD Developer Cloud), validate model fit, apply the right vLLM recipe, and launch a benchmarked endpoint. SGLang and engine or backend selection in later phases.	in-repo
serving-llms-on-epyc	Serve LLMs on AMD EPYC CPUs with vLLM and zentorch, in a container (Docker or Podman) or conda. Handles CPU detection, runtime and env validation, vLLM model-support and RAM-fit checks, hardware-sized threads and KV, launch, and health verification. Single instance; reports and stops on failure.	ZenDNN-pytorch-plugin
hyperloom-workload-optimizer	Set up Hyperloom and autonomously optimize end-to-end LLM inference throughput on AMD Instinct GPUs, reporting a validated gain.	Hyperloom
magpie-kernel-evaluator	Evaluate GPU kernel correctness and performance, compare kernel implementations, and benchmark vLLM or SGLang inference with profiling, TraceLens, and torch-trace gap analysis.	Magpie
tracelens-analysis-orchestrator	Orchestrate modular PyTorch profiler trace analysis with TraceLens: generate perf reports, run system-level and compute-kernel subagents in parallel, and write a prioritized stakeholder report.	TraceLens
rocprofiler-pc-sampling	Profile a GPU application with rocprofv3 program counter (PC) sampling, then analyze the samples to surface hotspots and stall reasons.	rocprofiler-sdk
rocprofiler-systems-choosing-a-profiling-preset	Pick the right rocprofiler-systems profiling preset for a workload and goal (GPU kernels, MPI, OpenMP, AI/ML training, and more), and return a ready-to-run command.	rocprofiler-systems
rocprofiler-systems-discovering-profiling-options	Discover rocprofiler-systems settings, hardware counters, ROCm API tracing domains, and GPU metrics, and how to configure them.	rocprofiler-systems
rocprofiler-systems-instrumenting-binaries	Instrument programs and libraries with rocprof-sys-instrument: choose runtime instrumentation or binary rewrite, filter functions and modules, preview with a dry run, and debug missing or noisy trace entries.	rocprofiler-systems
flydsl-kernel-authoring	Write, review, and understand FlyDSL GPU kernels for AMD GPUs: layout algebra, tiled copy and MMA, LDS allocation, and autotuning.	FlyDSL
flydsl-kernel-trace-analysis	Collect rocprofv3 instruction-level traces and counters for a GPU kernel, map the top stall hotspots back to source lines, and produce an optimization plan.	FlyDSL