Toolbox Setup
Build and configure the halo-forge container environment
The halo-forge toolbox is a containerized environment with ROCm and PyTorch configured for Strix Halo.
Why a Toolbox?
AMD Strix Halo (gfx1151) requires:
- ROCm 7 nightly builds (not yet in stable releases)
- PyTorch compiled specifically for gfx1151
- Specific library versions for unified memory
The toolbox packages all of this into a reproducible environment.
Building
cd halo-forge/toolbox
./build.sh
Build time: ~15-30 minutes depending on network speed.
What Gets Installed
| Component | Version | Notes |
|---|---|---|
| ROCm | 7.x nightly | From TheRock project |
| PyTorch | 2.6 nightly | Built for gfx1151 |
| Transformers | 4.46+ | HuggingFace |
| PEFT | 0.13+ | LoRA adapters |
| TRL | 0.12+ | Training utilities |
Usage
# Create the toolbox
toolbox create halo-forge --image localhost/halo-forge:latest
# Enter toolbox
toolbox enter halo-forge
# Check GPU access
python3 -c "import torch; print(torch.cuda.is_available())"
# Exit
exit
Persistence
Your home directory is automatically mounted in the toolbox:
~/on host =~/in toolbox- Models, datasets, checkpoints all persist
- HuggingFace cache at
~/.cache/huggingface/
Build Options
# Build with cache (default, faster)
./build.sh
# Build from scratch (slower, ensures fresh)
./build.sh --no-cache
Updating
cd halo-forge/toolbox
git pull
./build.sh --no-cache
Troubleshooting
Build Failures
Ensure you have:
- Sufficient disk space (~20GB for images)
- Network access to github.com, huggingface.co, PyPI
GPU Not Detected in Toolbox
Check device permissions on the host:
ls -l /dev/dri /dev/kfd
The toolbox automatically mounts these devices, but permissions must allow access.