The First Open-Source Agent Skills Collection for AMD ROCm

While NVIDIA-focused agent skills exceed 428 entries on the skills.sh registry used by AI coding agents like Claude Code, Cursor, and OpenCode, AMD ROCm had exactly zero. To close this gap, an open-source project called amd-rocm-skills was built, delivering 10 production-ready skills covering ROCm setup, Docker GPU passthrough, vLLM and YOLO deployment, video pipelines, benchmarking, and industrial PPE detection.

The technical foundation relies on the fact that PyTorch's torch.cuda API works identically on both AMD ROCm and NVIDIA CUDA, with torch.version.hip used to distinguish backends automatically. Every skill auto-detects ROCm, CUDA, or CPU and runs the same codebase across all three, while a single docker-compose file switches between rocm, nvidia, and cpu profiles.

Because every skill follows the agentskills.io specification without any agent-specific fields, they work identically across nine or more coding agents, including Claude Code, OpenCode, Cursor, Codex, Cline, Roo Code, Windsurf, Gemini CLI, and Kiro CLI. This lets engineers automate GPU setup, model deployment, and industrial vision tasks on AMD hardware through natural-language instructions instead of manual documentation lookups.

Installable with a single npx skills add command, the toolkit supports hardware ranging from data-center accelerators like the MI300X to consumer cards like the RX 7900, filling a significant gap in AMD's presence within the emerging agent-driven AI development ecosystem.