Installing ROCm

gfx1030 (Navi 21) is on AMD’s officially supported list for recent ROCm releases on Linux, so in most cases a stock install “just works” — no override hacks required.

Always follow the official ROCm install guide for your exact distro and ROCm version. The commands below are a convenience summary and may drift over time.

Supported operating systems

ROCm’s RDNA2 support targets specific LTS releases. As of recent ROCm versions that typically means:

Ubuntu 22.04 / 24.04 LTS
RHEL / Rocky 9.x
Debian 12 (community, less tested)
Fedora 43 (wiki-validated for power / P2P); Fedora Server 44 + kernel 6.19 reported working for the V620 powerfix (community)

Check the system requirements page for the version you plan to install.

Ubuntu quick install (amdgpu-install)

# 1. Add the amdgpu package repository (replace VERSION with the ROCm release you want, e.g. 6.4.60400-1)
sudo apt update
wget https://repo.radeon.com/amdgpu-install/latest/ubuntu/jammy/amdgpu-install_VERSION_all.deb
sudo apt install ./amdgpu-install_VERSION_all.deb
sudo apt update

# 2. Install ROCm (compute use case)
sudo amdgpu-install --usecase=rocm

# 3. Add yourself to the render/video groups so you can access the GPU without root
sudo usermod -aG render,video "$LOGNAME"

# 4. Reboot
sudo reboot

Verify the install

rocminfo                       # should list your Navi 21 card and "Name: gfx1030"
clinfo | grep -i 'Board\|gfx'  # OpenCL view
rocm-smi                       # live clocks, temps, VRAM, power

If rocminfo shows Name: gfx1030, you’re done. If it shows gfx1031/gfx1032/etc., your card is a smaller RDNA2 die — continue to HSA_OVERRIDE for RDNA2 Cousins.

Multi-GPU: pin ROCm 7.2.0 or 7.14.0

For more than one card, stay on ROCm 7.2.0 (not 7.2.1+) or jump to 7.14.0. #vllm-rdna / #llamacpp report an RCCL bug from 7.2.1 upward (including mid-line builds such as 7.2.4) that shows up as soon as you leave a single GPU — tensor-parallel hangs, comm failures, AllReduce init failed, or cards dropping offline. The published vllm-rdna images already sit on those two bases for that reason.

Do not “upgrade within 7.2.x” on a multi-GPU box. If you are already on a broken 7.2.1–7.13 userspace, rebuild or pull a 7.2.0 or 7.14.0 image rather than debugging RCCL on the in-between releases. Ubuntu 26.04 can ship ROCm/HIP 7.1 (#llamacpp Sep 17) — treat that as an A/B, not the multi-GPU pin. ROCm 10 / TheRock packaging is interesting for newer cards; community llama.cpp benches on Ice Lake have run ROCm 10.0 with numbers close to 7.1 on that host. #vllm-rdna (29 Oct): a fresh Ubuntu 26.04 install can land /opt/rocm/core-10 (and core-10.1) with no core-7.14 folder — do not copy a 7.14 ROCM_HOME from host venv onto that tree. Same thread: ROCm 10 serves on at least one host; the Triton wheel for that userspace is unconfirmed. #vllm-rdna (30 Oct): maintainer called ROCm 10 and 7.14 the same for day-to-day serving; another host is already serving on 10 with no issues so far. That is not a reason to wipe a working 7.14 box. One 10.1 host-venv attempt failed before a clean 7.14 retry — Needs verify, not a 10.1 ban. Day-to-day gfx1030 multi-GPU advice is still: prefer 7.14.0 (or 7.2.0) unless you are deliberately lab-testing 10.x.

Notes & gotchas

Kernel driver: ROCm relies on the amdgpu kernel module. Very new kernels sometimes ship a driver newer than your ROCm userspace expects; the amdgpu-dkms package from the amdgpu repo keeps them in sync.
Secure Boot: if Secure Boot is enabled, the DKMS module must be signed or it will fail to load. The easiest path for a dev box is to disable Secure Boot.
Multiple GPUs / iGPU present: if your CPU also has an RDNA2 iGPU, ROCm may enumerate it. Pin the discrete card with HIP_VISIBLE_DEVICES (see Reference).
hipBLASLt: some libraries assume hipBLASLt, which historically had gaps on Navi 21. If a workload complains, try TORCH_BLAS_PREFER_HIPBLASLT=0 (PyTorch) or the workload’s equivalent flag.