softab

Community Resources and References

Active community maintaining Strix Halo AI software stack

Primary Documentation

Resource	URL	Focus
Strix Halo Wiki	https://strixhalo.wiki	Comprehensive setup guides
lhl’s testing repo	https://github.com/lhl/strix-halo-testing	Benchmarks, build scripts
llm-tracker.info	https://llm-tracker.info/_TOORG/Strix-Halo	Deep technical documentation
kyuz0 toolboxes	https://github.com/kyuz0/amd-strix-halo-toolboxes	Docker containers

Interactive Tools

• Benchmark viewer: https://kyuz0.github.io/amd-strix-halo-toolboxes/
• Framework forums: https://community.frame.work (search “Strix Halo”)
• Discord: https://discord.gg/pnPRyucNrG

Key Repositories

Repository	Purpose	Best For
kyuz0/amd-strix-halo-toolboxes	llama.cpp containers	LLM inference
kyuz0/amd-strix-halo-vllm-toolboxes	vLLM containers	Production serving
scottt/rocm-TheRock	PyTorch wheels	CV/ML research
lhl/strix-halo-testing	Comprehensive test scripts	Benchmarking
llm-tracker.info	Ongoing performance notes	Latest findings
lemonade-sdk/llamacpp-rocm	Pre-built llama.cpp binaries	Quick setup
hjc4869/llama.cpp	Optimized fork	David Huang’s optimizations
nix-strix-halo	Nix flake	Reproducible builds

For Computer Vision/PyTorch work: Use scottt/rocm-TheRock path. The kyuz0 llama.cpp containers are LLM-focused and won’t help for EfficientLoFTR, diffusion models, etc.

Lemonade Server (AMD Official)

Resource	URL
Main site	https://lemonade-server.ai/
FAQ/Docs	https://lemonade-server.ai/docs/faq/
GitHub	https://github.com/lemonade-sdk/lemonade
ROCm llama.cpp	https://github.com/lemonade-sdk/llamacpp-rocm
AMD Tutorial	https://www.amd.com/en/developer/resources/technical-articles/2025/ryzen-ai-radeon-llms-with-lemonade.html

Additional Resources

Resource	URL
Ryzen AI SDK docs	https://ryzenai.docs.amd.com/en/latest/relnotes.html
Level1Techs benchmarks	https://forum.level1techs.com/t/strix-halo-ryzen-ai-max-395-llm-benchmark-results/233796
ROCm Issue Tracker	https://github.com/ROCm/ROCm/issues
AMD Developer Forums	https://community.amd.com/t5/ai/ct-p/amd-ai

Kernel and Container Considerations

Understanding Container Architecture

Critical: Containers are NOT virtual machines. They share the host kernel.

Component	Location	Can Vary in Container?
Kernel	Host	❌ NO - shared by all containers
Kernel Modules (amdgpu, etc.)	Host	❌ NO - loaded in host kernel
Device Files (/dev/kfd, /dev/dri)	Host	❌ NO - bind-mounted from host
ROCm Userspace (rocBLAS, hipBLAS)	Container	✅ YES - part of container image
Python Version	Container	✅ YES - isolated per container
Application Binaries	Container	✅ YES - isolated per container

Why Kernel Matters for GPU Workloads

The host kernel directly impacts GPU performance and compatibility:

Kernel Feature	Impact on Strix Halo
amdgpu driver version	Determines GPU support, bug fixes
VRAM visibility	6.16.9+ fixes bug showing only ~15.5GB
Kernel ABI	6.18.4+ may need matching ROCm userspace
Memory management	GTT/GART configuration via kernel params

Container vs Host Components

Container Image Contains:

• ✅ ROCm libraries (rocBLAS, hipBLASLt, MIOpen)
• ✅ Python interpreter and packages (PyTorch, etc.)
• ✅ Application binaries (llama.cpp, whisper.cpp)
• ✅ Environment variables (HSA_OVERRIDE_GFX_VERSION, etc.)

Host Provides:

• ✅ Linux kernel (all containers use same kernel)
• ✅ amdgpu kernel module (GPU driver)
• ✅ Device access (/dev/kfd, /dev/dri)
• ✅ Firmware (GPU microcode)

Kernel Ablation: Practical Limitations

To test different kernel versions, you must:

1. Install multiple kernels on the host
2. Reboot the host to switch kernels
3. Re-run container tests on the new kernel
4. Cannot test kernels in parallel (single host, one kernel at a time)

Recommendation for ablation studies: Treat kernel as a constant, not a variable. Testing kernels requires host reboots and serialized testing, which is expensive and out of scope for software stack configuration testing.

Checking Your Kernel

# Check current kernel
uname -r

# List installed kernels
rpm -qa kernel | sort -V        # Fedora
dpkg -l | grep linux-image      # Ubuntu

# List boot entries
sudo grubby --info=ALL | grep -E "(title|kernel)"

# Set kernel for next boot
sudo grubby --set-default /boot/vmlinuz-VERSION

# Reboot
sudo reboot

Glossary

Term	Definition
gfx1151	AMD GPU architecture identifier for Strix Halo
RDNA 3.5	GPU architecture (between RDNA 3 and RDNA 4)
GTT	Graphics Translation Table - dynamic GPU memory
GART	Graphics Address Remapping Table - fixed GPU memory
rocWMMA	ROCm Wave Matrix Multiply-Accumulate library
hipBLASLt	AMD’s lightweight BLAS library for HIP
AOTriton	AMD’s OpenAI Triton fork
TheRock	AMD’s ROCm nightly build system
pp/tg	Prompt processing / token generation (llama.cpp metrics)
FA	Flash Attention
Toolbox	Fedora’s containerized development environment
Distrobox	Universal container manager supporting multiple distros
VRAM estimator	Tool to calculate GPU memory needs for GGUF models
RPC	Remote Procedure Call - enables distributed inference
MES	Micro Engine Scheduler (GPU job scheduler)
SDMA	System DMA engine (can cause artifacts if enabled)
HSA	Heterogeneous System Architecture (AMD’s programming model)

Version History

• 2026-01-26: Split KNOWLEDGE_BASE.md into topic-specific documentation
  • Created docs/ directory structure
  • Organized by hardware, ROCm, applications, troubleshooting, community
• 2026-01-24: Added comprehensive kyuz0 toolboxes documentation
  • Pre-built Vulkan and ROCm containers for llama.cpp
  • Critical runtime flags (–no-mmap, -fa 1) documented
  • VRAM estimator and distributed inference tools
  • Updated kernel parameters with kyuz0’s tested configuration
• 2026-01: Initial comprehensive documentation
  • Based on community work through January 2026
  • Software stack rapidly evolving - verify current status before major decisions

---

Back to: KNOWLEDGE_BASE.md