amd-skills Claude Code Plugin
Claude Code PluginClaude CodeDevelopmentAI Agents & AI App DevelopmentClaude Code Customization & WorkflowCode Quality, Review & TestingThe explanation below is AI-generated. Please verify it against the sources.
amd-skills is a plugin that packages a catalog of Agent Skills for building and optimizing AI workloads on AMD hardware and software. According to its README, the catalog spans client-native skills for routing image, text-to-speech, and speech-to-text tasks through a local AI server on Ryzen AI, cross-stack skills such as a ROCm diagnostics helper and a Lemonade model router, and server-native skills for deploying LLM inference on AMD Instinct GPUs and EPYC CPUs with vLLM. It also includes skills for tuning inference throughput with Hyperloom, evaluating GPU kernel correctness and performance with Magpie, and orchestrating PyTorch profiler trace analysis with TraceLens. The README documents installation via an npx-based skills CLI or manual copying of skill folders, and the project is released under the MIT License.
What you can do with amd-skills
- Integrate local AI into cloud LLM applications for offline use, added privacy, and lower API costs cpp failures on AMD GPUs (listed as planned in the README)
- Configure a Lemonade model router that directs requests by content, sensitivity, or capability
- Record, replay, and analyze GPU workload behavior via HIP Record and Replay archives (listed as planned)
- Deploy end-to-end LLM inference on AMD Instinct GPUs using vLLM recipes with hardware detection and benchmarking
- Evaluate and compare GPU kernel correctness and performance, including vLLM/SGLang benchmarking, using Magpie
- Orchestrate modular PyTorch profiler trace analysis and generate prioritized performance reports using TraceLens
Sources
History of amd-skills
- Claude Code Plugin Added amd-skills