Getting Started

This page gets you from a bare Linux install to a working ROCm + PyTorch stack on a gfx1030 card, and explains how to preview this wiki locally.

1. Confirm your hardware

Make sure your card is a Navi 21 gfx1030 (or a related RDNA2 card):

lspci | grep -i vga
# After ROCm is installed:
rocminfo | grep -m1 -o 'gfx[0-9]*'

See Supported Hardware for the full card list. If the box never POSTs with several V620s (especially behind a PLX switch), read Host firmware before chasing ROCm.

2. Install ROCm

Follow Installing ROCm. At a high level:

# Add the amdgpu repo, then:
sudo apt install rocm
sudo usermod -aG render,video "$LOGNAME"
# Reboot, then verify:
rocminfo
clinfo | grep -i 'gfx\|Board'

3. (Optional) Tune the card

If you run one or more Radeon PRO V620 (or PRO W6800) cards, these tweaks are worth doing before you load models (Fedora or Ubuntu 26.04 for power cap; P2P is Fedora + AMD-validated — A/B on Intel, see When P2P helps):

• Power Tuning — drop the VBIOS-locked 250 W floor to 120 W and boot-cap at 180 W. Passthrough VM and need OD clocks instead? See Soft unlock (amd-v620-soft-unlock).
• Disabling ECC — Pro cards hide ~2 GB behind ECC; optional extra VRAM.
• Multi-GPU PCIe P2P — enable GPU↔GPU peer-to-peer for multi-card setups.

Baremetal power floor + P2P come from the v620_toolbox repo.

4. Run an inference stack (Docker)

The fastest path is the prebuilt images — no local ROCm/PyTorch/vLLM build required:

• Onboard vLLM on V620 — short happy path for Hub v0.31.0-extras or the fork launcher (serve_rdna.sh).
• Running vLLM (Docker) — blivioniag/vllm-rdna on a blivioniag/rocm-rdna PyTorch base.
• vLLM recipes — Hub -extras vs recipe container vs Flash-Next by card count.

Prefer GGUF and building from source? See Building & Running llama.cpp (ROCm or Vulkan).

Want to build the images yourself, or use the RDNA-tuned kernels? See Building the Images and The rdna_extras fork.

Need the same stack in a host uv venv instead of Docker? See Host venv (ROCm 7.2.0 index wheels, or 7.14.0 multi-arch wheels).

5. Smoke test

# Inside a rocm-rdna / vllm-rdna container, or a local ROCm PyTorch env:
python -c "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0))"

You should see True and your Radeon card’s name. If not, see Troubleshooting.

6. Join the community Discord

Questions, live benches, and fork updates land first on gfx1030 club: https://discord.gg/mESex2aBp (#vllm-rdna, #llamacpp, #general, #benchmarks). See Useful resources.

Editors: local preview and page conventions are in Contributing.