softab

Strix Halo Software Stack - Quick Reference

Scope

This project tests software stack compatibility and performance on Strix Halo:

• GFX target, driver (AMDVLK/RADV), ROCm version, installation method, backend
• NOT model comparisons, quantization sweeps, or inference tuning

What Works

Stack	PyTorch	whisper.cpp	llama.cpp
TheRock nightlies + gfx1151	✅	✅	✅
scottt/rocm-TheRock wheels	✅	N/A	N/A
AMD ROCm 6.4.4 + HSA_OVERRIDE	⚠️ Unstable	✅	✅
Fedora ROCm packages	❓	✅	✅
Official PyTorch wheels (rocm6.2)	❌	N/A	N/A
Vulkan backend	N/A	N/A	✅

Required Environment Variables

# Always needed for stability
export HSA_ENABLE_SDMA=0

# For PyTorch
export PYTORCH_HIP_ALLOC_CONF="backend:native,expandable_segments:True"

# If using HSA_OVERRIDE workaround (not recommended)
export HSA_OVERRIDE_GFX_VERSION=11.0.0

Kernel Parameters

Add to GRUB_CMDLINE_LINUX_DEFAULT:

amd_iommu=off amdgpu.gttsize=131072 amdgpu.cwsr_enable=0

llama.cpp Flags

--no-mmap -ngl 999    # Required for ROCm backend

Common Failures

Symptom	Cause	Fix
“invalid device function”	Missing gfx1151 kernels	Use TheRock nightlies
Only 15.5GB VRAM visible	Kernel bug	Upgrade to 6.16.9+
MES/kernel hangs	CWSR issue	Add amdgpu.cwsr_enable=0
Checkerboard artifacts	SDMA bug	Set HSA_ENABLE_SDMA=0
Build fails on kernel 6.18+	ROCm incompatible	Use TheRock nightlies

Installation Paths

Fedora 43 (Recommended)

sudo dnf install rocm-hip-devel

TheRock Nightlies (for kernel 6.18+)

# See docker/rocm/Dockerfile.therock

PyTorch with gfx1151 Support

# Option 1: scottt wheels (recommended)
pip install "https://github.com/scottt/rocm-TheRock/releases/download/v6.5.0rc-pytorch/torch-2.7.0a0+gitbfd8155-cp311-cp311-linux_x86_64.whl"

# Option 2: AMD nightlies
pip install --index-url https://rocm.nightlies.amd.com/v2/gfx1151/ --pre torch

Key Resources

• Wiki: strixhalo.wiki
• Testing repo: github.com/lhl/strix-halo-testing
• Containers: github.com/kyuz0/amd-strix-halo-toolboxes