Host Configuration Guide for AMD Strix Halo

This document details host-level configuration for running the Gufo Docker and Podman images reliably on AMD Ryzen AI Max ("Strix Halo", gfx1151 GPU + XDNA2 NPU).

1. Kernel and Firmware Requirements

Recommended Setup

Linux Kernel: >= 6.18.4 (e.g. 6.18.9+). Kernels older than 6.18.4 exhibit stability bugs and crashes under sustained ROCm compute on gfx1151.
Linux Firmware: 20260110 or later.
Avoid: linux-firmware-20251125 (breaks ROCm initialization on gfx1151).

Kernel Boot Parameters

Add the following parameters to your bootloader (e.g. GRUB or systemd-boot):

amdgpu.gttsize=65536

(On 128 GiB unified memory systems, this ensures sufficient GTT allocation for large model weights and KV caches).

2. Device Permissions and User Groups

Ensure your host user account belongs to the required hardware access groups:

sudo usermod -aG video,render $USER

Verify device existence:

# GPU ROCm / KFD interface
ls -l /dev/kfd /dev/dri/renderD*

# XDNA2 NPU interface
ls -l /dev/accel/accel* /dev/amdxdna 2>/dev/null || echo "NPU driver not loaded"

3. Ulimits (Locked Memory)

Large unified memory buffers require unlimited memlock limits. In /etc/security/limits.d/99-gufo.conf:

* soft memlock unlimited
* hard memlock unlimited

4. Power Profiles

For sustained LLM inference on Strix Halo laptops or mini PCs, switch power profiles to throughput-performance:

sudo tuned-adm profile throughput-performance

(Or install the provided systemd/gpu-workload-watch service to automate this on GPU activity).

1.7 KiB Raw Permalink Blame History