Install ROCm on a home Linux box

This walkthrough is for a dedicated or dual-boot Linux install you are willing to reboot. It targets the AMD Radeon AI PRO R9700 (gfx1201) and other GPUs on AMD’s current matrix. It is not a silent unattended script. For what ROCm is — stack, gfx IDs, Vulkan vs HIP, versions — start at What is ROCm.

What 10.1.0 actually lists

From the ROCm 10.1.0 release notes (as of that page):

• Radeon AI PRO R9700 is on the hardware table as gfx1201 (RDNA 4), with the RX 9070 / 9070 XT.
• Radeon OS rows: Ubuntu 26.04.1 (GA 7.0), 24.04.5 (HWE 7.0), 22.04.5 (HWE 6.8). Instinct has its own kernel footnotes — do not mix those rows.
• 10.1.0 adds Radeon AI PRO R9600; the R9700 is already on the 10.0/10.1 device tables.

The 10.1 blog (5 Oct 2026) is the narrative: Infinity Storage / hipFILE, NUMA-aware HIP alloc, LLVM 24, amd-smi succeeding rocm-smi, WSL2 tech preview. Re-read both on install day.

What you will have at the end

• Kernel sees the GPU
• A single ROCm userspace (AMD repo or CLI-managed or runfile — not a pile)
• Your user can talk to /dev/kfd and /dev/dri
• rocminfo prints the marketing name and gfx1201 (or your card’s target)

Prerequisites

• A 64-bit x86 PC with the R9700 (or another listed card) in a PCIe slot, power cable seated, and a display or SSH path that still works if the GUI driver hiccups.
• An OS named on the matrix for that GPU. For home Radeon, 10.1.0 names Ubuntu 26.04.1, 24.04.5, and 22.04.5. Ubuntu LTS is the default this tutorial assumes.
• Disk space for ROCm plus later models (20 GB free is a low floor; more is kinder).
• Time to reboot twice if the first driver pass is messy.

Decision tree — do not mix apt sources

1. Already have Ubuntu archive rocm / libamdhip64-dev / libhsa-runtime*? Purge those first.
2. Pick one method: (A) AMD 10.1 repo + one gfx family, (B) ROCm CLI v0.2.0 rocm install sdk, (C) amdgpu-install usecases, or (D) runfile / tarball.
3. Never add Canonical rocm on top of an AMD 10.x tree.

Why purge first. ROCm #6110: gfx1201 “has 2 ISAs but can only support a single ISA” was stale Ubuntu HIP/HSA libraries shadowing a newer /opt/rocm. Archive apt install rocm has lagged (7.1-era reports on new Ubuntu). It will not give you 10.1, amdrocm-*, or a current vLLM ROCm wheel.

# Inventory — do not copy a purge list from memory if you do not recognize the packages
dpkg -l | grep -Ei 'rocm|amdrocm|hip|hsa-runtime|amdgpu-install'

Remove the archive rocm / libamdhip64-dev / libhsa-runtime* set if you are about to add AMD’s 10.1 repo. Then pick one method below.

1. Confirm the hardware before you add repos

uname -r
. /etc/os-release && echo "$PRETTY_NAME"
lspci -nn | grep -Ei 'vga|display|3d'

You want a line that looks like an AMD device. Note the PCI ID if you file bugs later.

lsmod | grep amdgpu

If amdgpu is already loaded from the distro kernel, that is normal. Decide now: keep the distro kernel module and add userspace, or install AMD’s DKMS package as the docs for your chosen method require. Do not do both blindly.

2. Open the official matrix and pick a ROCm version

1. 10.1.0 release notes — GPU row + Radeon OS row.
2. Install ROCm 10.1.0 — method selector.
3. Compatibility matrix.
4. If you specifically want Radeon desktop graphics + ROCm together, also skim ROCm on Radeon and Ryzen.

Write down: ROCm version, OS codename, install method. One method only.

3. Install using one official method

Option A — AMD package manager (10.1 / TheRock)

Follow the Ubuntu chapter on Install ROCm 10.1.0. The shape (names move — copy from the docs):

# 1) Register AMD's signed 10.1 repo — copy the exact lines from the docs
#    for your Ubuntu point release (26.04.1 / 24.04.5 / 22.04.5).
sudo apt update

# 2) See what family packages exist THIS week
apt-cache search amdrocm-core-sdk

# 3) Install ONE gfx family. R9700 is RDNA4 gfx1201.
#    Example shape only — confirm the suffix:
# sudo apt install amdrocm-core-sdk10.1-gfx1201

Confirm the prefix (ls /opt/rocm) and put bin on PATH. Installing every family package at once is how you fight TheRock. 10.0 prefixes were often /opt/rocm/core-10.0. Do not invent a 10.1 prefix — ls it.

Option B — ROCm CLI (published v0.2.0)

The 10.1 blog calls this ROCm CLI v0.2.0. GitHub’s latest stable tag on 2026-10-07 is v0.2.0. There is no v1.0.0 tag as of 2026-10-12 — if someone told you 1.0.0, confirm with rocm version after install.

Getting started:

curl -fsSL https://raw.githubusercontent.com/ROCm/rocm-cli/main/install.sh | sh
rocm version
rocm examine
rocm install sdk

rocm install sdk downloads TheRock wheels into a managed environment. rocm examine will call an existing /opt/rocm legacy_rocm_status: detected_unmanaged. That is fine — do not then also apt install rocm from Ubuntu. The CLI can install vLLM 10.1 wheels (rocm install vllm); that does not make vLLM the consumer default. Backends.

Option C — amdgpu-install usecases

# Download the amdgpu-install package for your Ubuntu release from the
# versioned docs, then:
sudo amdgpu-install --list-usecase
sudo amdgpu-install -y --usecase=graphics,rocm

--usecase=rocm without graphics is a thing on headless boxes. --no-dkms is a thing if you intentionally keep the distro kernel module. Only use flags you can explain.

Option D — Runfile or tarball

Runfile installer when the machine cannot grow extra apt repos. It still needs a compatible kernel driver underneath. Copy the current .run name from that page — do not invent a filename.

4. Groups, paths, and a reboot

sudo usermod -aG render,video "$USER"

Log out completely or reboot so the groups apply. New shells are not enough if you never left the session. Permissions pattern.

If the docs tell you to add ROCm bin to PATH, do that in ~/.profile after ls /opt/rocm.

sudo reboot

5. Minimum verification (then stop)

groups | grep -E 'render|video'
which rocminfo rocm-smi amd-smi
rocminfo | less
rocm-smi
# 10.1 blog: rocm-smi deprecation period is over; amd-smi is the successor

In rocminfo, find the agent that is a GPU, not the CPU. You want a marketing name that matches the card and a GFX target of gfx1201 on the R9700.

If rocminfo is missing, the userspace packages did not land. If it runs but lists only the CPU, the KFD device is not usable — groups, driver, or a GPU the userspace cannot target.

Do not install PyTorch yet if this step failed. Continue in Verify GPU and ROCm.

Common failure modes

Black screen after reboot. You mixed a desktop amdgpu stack with another driver story, or DKMS failed against this kernel. Recovery: TTY (Ctrl+Alt+F3) or SSH, then the official uninstall / purge steps for the method you used. Do not apt install a third driver to “fix” it.

Permission denied on /dev/kfd. Groups did not apply, or you are inside a container without the right devices.

Card missing from rocminfo. Wrong ROCm for gfx1201, mixed archive + AMD libs, IOMMU / ACS oddities, or the card never initialized (dmesg | grep -i amdgpu).

Secure Boot. DKMS modules will not load unsigned.

MODE1 reset / MES hang. Paste the dmesg.

What not to do

• Do not install CUDA-shaped advice (nvidia-smi, nvcc) and expect a translation layer.
• Do not pin a blog’s ROCm 6.1 commands onto a 2026 R9700 and then override GFX versions.
• Do not sudo every inference process to dodge group permissions.
• Do not run Ubuntu archive rocm and AMD’s 10.1 repo at the same time.

When rocm-smi / amd-smi looks sane, run a local LLM.