Strix Halo ROCm Quick Start¶
Single-page path from bare Ubuntu 26.04 LTS to running LLMs on GPU with the AMD Ryzen AI Max+ 395 APU (gfx1151).
Tested Environment
Ubuntu 26.04 LTS ("Resolute Raccoon"), Linux 7.0 kernel, ROCm 7.1 (in-distro) or upstream 7.2.x, 128GB LPDDR5X-8000 quad-channel. For detailed configuration of individual components, see the linked pages throughout this guide.
1. Verify Kernel¶
Ubuntu 26.04 ships Linux 7.0 by default and the server installer auto-installs HWE/OEM metapackages where applicable, so no manual kernel install is needed for gfx1151. Confirm:
AMD's official requirement for Ryzen AI Max (gfx1151) is kernel 6.18.4+ (mainline) or 6.17 HWE — Ubuntu 26.04's 7.0 kernel clears this.
2. Install ROCm 7.x¶
Option A: in-distro ROCm 7.1 (one command, fewest moving parts)
Ubuntu 26.04 ships ROCm 7.1.0 in Universe. This is roughly one minor behind upstream's 7.2.x but receives Ubuntu security updates and avoids third-party repos.
Option B: upstream AMD repo (newer ROCm, faster cadence)
# AMD's resolute path may not be live yet at release; verify on repo.radeon.com first.
wget https://repo.radeon.com/amdgpu-install/latest/ubuntu/noble/amdgpu-install_7.1.1.70101-1_all.deb
sudo apt install ./amdgpu-install_7.1.1.70101-1_all.deb
sudo amdgpu-install --usecase=rocm,hip,opencl,graphics,dkms
Add your user to the required groups (either option):
3. Verify GPU Detection¶
rocminfo | grep gfx
# Should show: Name: gfx1151
rocm-smi
# Should show the AMD Radeon device with temperature, power, etc.
No HSA Override Needed
ROCm 7.x has native gfx1151 support. You do not need HSA_OVERRIDE_GFX_VERSION.
4. Allocate VRAM with amd-ttm¶
By default, the system allocates roughly 62GB as GPU-accessible memory. For large LLMs, allocate more using amd-debug-tools:
# Install amd-debug-tools
pipx install amd-debug-tools
# Check current allocation
amd-ttm
# Shows current GTT (Graphics Translation Table) size
# Set to 108GB (leaves ~20GB for OS)
amd-ttm --set 108
sudo reboot
Verify after reboot:
amd-ttm
# Should confirm ~108GB allocation
rocm-smi --showmeminfo vram
# Should reflect the new allocation
For more details on memory allocation strategies, see Memory Configuration.
5. Install and Test Ollama¶
Run a test model:
Verify GPU usage:
6. llama.cpp (Docker)¶
For llama.cpp via Docker with ROCm:
docker run -d \
--device=/dev/kfd \
--device=/dev/dri \
--group-add video \
--group-add render \
-v /path/to/models:/models \
-p 8080:8080 \
ghcr.io/ggml-org/llama.cpp:server-rocm \
-m /models/your-model.gguf \
--host 0.0.0.0 \
-ngl 99
See GPU Containers for Docker Compose examples.
Troubleshooting¶
GPU not detected by rocminfo¶
# Check amdgpu driver is loaded
lsmod | grep amdgpu
# If not loaded
sudo modprobe amdgpu
# Check for blacklist entries
grep -r amdgpu /etc/modprobe.d/
# Remove any blacklist lines if found
# Rebuild initramfs after changes
sudo update-initramfs -u
sudo reboot
Wrong kernel booting¶
# Check which kernels are installed
dpkg -l | grep linux-image
# Set the desired kernel as default (replace 7.0.x-N-generic with the actual entry)
sudo sed -i 's/GRUB_DEFAULT=.*/GRUB_DEFAULT="Advanced options for Ubuntu>Ubuntu, with Linux 7.0.0-N-generic"/' /etc/default/grub
sudo update-grub
sudo reboot
rocm-smi shows no devices¶
Ensure your user is in the video and render groups:
groups $USER
# Should include: video render
# If not, add and re-login
sudo usermod -aG video,render $USER
# Log out and back in
What's Next¶
- ROCm Installation -- detailed installation options and environment configuration
- Memory Configuration -- VRAM allocation strategies and kernel parameters
- Driver Updates -- keeping ROCm current
Sources¶
This guide consolidates information from the following community resources:
- Shoresh613 - ROCm on Strix Halo -- comprehensive setup notes
- hakedev - ROCm for Strix Halo -- kernel and driver walkthrough
- pablo-ross - Strix Halo ROCm setup -- tested configuration
- Jeff Geerling - AMD Strix Halo -- hardware and software notes