stxh-setup-amd 1.0.0

Automated setup for AMD Strix Halo APU: ROCm, MIGraphX, custom ORT, and AI model deployment.

stxh-setup-amd

Automated setup for AMD Strix Halo APU (gfx1151 / RDNA 3.5).

Takes a fresh Ubuntu 24.04 machine from zero to running 23 AI models on ORT MIGraphX Execution Provider -- ROCm 7.2, MIGraphX, 400 GB swap, GRUB tuning, Python ML stack, and a custom ONNX Runtime build with unified-memory fixes.

Install

pip install stxh-setup-amd

Usage

# Full setup (all 8 stages)
export SMB_USER="[e-mail]"
export SMB_PASS="password"
sudo -E stxh-setup --all

# Resume from a specific stage
sudo stxh-setup --from-stage 5

# Run a single stage
sudo stxh-setup --stage 3          # ROCm only

# Skip slow steps
sudo stxh-setup --all --skip-copy --skip-ort-build

# Check what's already done
sudo stxh-setup --status

# Verify everything works
sudo stxh-setup --verify

Fallback if stxh-setup is not on PATH:

sudo python3 -m stxh_setup --all

The 8 Stages

Stage	What it does	Time
1	System update + base packages (build-essential, cmake, ffmpeg, ...)	5 min
2	Copy R1models from network share (~206 GB)	30-90 min
3	GPU driver + ROCm 7.2 (requires reboot)	10 min
4	MIGraphX	2 min
5	Swap space (~400 GB for JIT compilation)	5 min
6	GRUB kernel parameters (amdgpu.gttsize=28672, requires reboot)	1 min
7	Python dependencies (PyTorch+ROCm, JAX+ROCm, TF, HuggingFace, ...)	15 min
8	Custom ORT 1.25.0 build with MIGraphX fixes (or installs existing wheel)	60-90 min

Each stage is idempotent -- if it detects the work is already done, it skips.

Custom ORT Fixes

The stock onnxruntime-migraphx from PyPI OOMs on any model > 512 MB because the Strix Halo APU only has 512 MB dedicated VRAM. Stage 8 applies two source patches before building:

Fix	What
B	hipMalloc -> hipMallocManaged -- spill to 28 GB GTT when VRAM is full
C	[=] -> [=, this] -- C++20 lambda capture build fix

Fix A (skip constant initializer double-allocation) is already upstream in ORT 1.25.0.

Hardware Target

• APU: AMD Strix Halo (gfx1151 / RDNA 3.5)
• VRAM: 512 MB dedicated + 28 GB GTT (unified memory)
• OS: Ubuntu 24.04 (Noble), kernel 6.18.0+
• ROCm: 7.2
• Key setting: HSA_XNACK=1 (demand paging for unified memory)

License

Apache 2.0

Copyright 2026 Sudheer Ibrahim Daniel Devu