rocmate

Get AMD GPUs running with AI tools — without the rabbit hole.

rocmate is a curated compatibility index and CLI for running modern AI workloads on AMD GPUs — via ROCm or Vulkan. It tells you what works on your specific card, what to set, and what to avoid — based on configurations that real users have tested.

pip install rocmate

Why?

AMD GPUs offer great VRAM-per-dollar (especially the RX 7900 XTX with 24 GB), but getting Ollama, ComfyUI, faster-whisper, or axolotl to actually use the GPU still involves hunting through blog posts, GitHub issues, and Discord threads. Information is scattered, often outdated, and rarely specific to your chip generation (gfx1030 vs gfx1100 vs gfx1151 vs gfx1201).

rocmate consolidates this knowledge into one place — version-controlled, testable, community-maintained.

Quickstart

pip install rocmate

Check your system:

$ rocmate doctor
✓ GPU detected: AMD Radeon RX 7900 XTX (gfx1100)
✓ ROCm found: 7.2.3
✗ User not in 'render' group — run: sudo usermod -aG render $USER
✓ Docker available
✓ Vulkan available

Show the tested config for a tool:

rocmate show llama-cpp               # auto-detects your chip
rocmate show llama-cpp --chip gfx1100

Find tools by keyword:

rocmate search llm
rocmate list --chip gfx1151          # filter by chip

Install with the right ENV vars (shows plan first, asks for confirmation):

rocmate install ollama
rocmate install ollama --chip gfx1100   # skip GPU detection

# Auto-fix detected issues
rocmate doctor --fix

Supported tools

→ Live compatibility matrix — auto-updated on every commit.

Tool	gfx1151 Radeon 8060S/8050S	gfx1100 RX 7900 XT/XTX	gfx1101 RX 7800/7700 XT	gfx1102 RX 7600	gfx1030 RX 6800/6900	gfx1201 RX 9070
Ollama	✅	✅	✅	✅	✅	🟡
ComfyUI	🟡	✅	✅	🟡	🟡	—
faster-whisper	✅	✅	🟡	—	—	—
llama.cpp	✅	✅	✅	✅	✅	🟡
Stable Diffusion WebUI	🟡	✅	✅	🟡	🟡	—
vLLM	✅	✅	✅	🟡	🟡	🟡
Axolotl	🟡	✅	✅	🟡	🟡	—
ExLlamaV2	🟡	✅	✅	🟡	🟡	—

✅ tested · 🟡 partial / workarounds needed · — no data yet

Run rocmate show <tool> for chip-specific ENV vars, install hints, and tested ROCm versions.

Status

Actively maintained by @T0nd3 on an RX 7900 XTX. Contributions for other AMD chips are very welcome — one YAML file, five minutes. See CONTRIBUTING.md.

License

MIT — see LICENSE.