Run a full AI stack locally on AMD — no NVIDIA, no CUDA

Self-hosted LLM chat, image generation, web search and agent tools on a single AMD Radeon GPU with ROCm. Private, offline-capable, and free of the NVIDIA tax.

View on GitHub → Quick start

ROCm gfx1201 RDNA4 No CUDA required rootless Podman MIT license

Local AI on AMD ROCm: Open WebUI, llama.cpp, ComfyUI and MCP tools running self-hosted on a Radeon GPU

Why run AI on AMD?

Large models usually mean paying the NVIDIA premium — but you don't have to. Modern AMD Radeon cards offer a lot of VRAM per dollar, and the open ROCm stack now runs the full local-AI toolchain well. ai-stack packages the images, flags and VRAM management so it works out of the box.

What you get

Quick start

git clone https://github.com/wildcommitter/ai-stack.git && cd ai-stack
./scripts/install.sh
./download-model.sh all <HF_TOKEN>
systemctl --user start ai-stack.service
# open http://localhost:3000

Guides

FAQ

Can I run a local LLM on an AMD GPU without NVIDIA or CUDA?

Yes. Every component runs on AMD ROCm — llama.cpp for the LLM, ComfyUI for images — no NVIDIA card or CUDA needed.

Is this a self-hosted, private ChatGPT alternative?

Yes. Open WebUI gives a familiar chat UI over a local model; nothing leaves your machine and there are no API keys or fees.

Does it do local image generation like Stable Diffusion?

Yes, via ComfyUI on ROCm, callable directly from the chat.

Which AMD GPU do I need?

Developed on the AMD Radeon AI PRO R9700 (32 GB, RDNA4, gfx1201). Other ROCm Radeon cards work with minor tweaks; more VRAM runs larger models.

How do the LLM and image model share limited VRAM?

They don't run at once — llama-swap unloads the chat model on idle and the image tool frees it before a ComfyUI job, so one 32 GB card handles both.