rocm-strix-docker

Docker infrastructure for AMD Strix Halo (RDNA 3.5 / gfx1151): a PyTorch + ROCm base container and a separate Ollama LLM service. Two folders, two Compose files, one Strix Halo box.

---

What this is

Two independent services, each in its own folder with its own docker-compose.yml:

• rocm/: base container with PyTorch + ROCm for ML workloads
• ollama/: Ollama LLM inference server with GPU acceleration

Project Structure

.
├── .gitignore                    # Ignores .env files (contain machine-specific values)
├── README.md                     # This file
├── llm.txt                       # Complete technical reference
├── rocm/                         # ROCm PyTorch container
│   ├── Dockerfile
│   ├── docker-compose.yml
│   └── entrypoint.sh
└── ollama/                       # Ollama LLM service
    ├── .env                      # Your local config (git-ignored)
    ├── .env.template             # Template to copy
    ├── Dockerfile
    ├── docker-compose.yml
    └── entrypoint.sh

Prerequisites

• Ubuntu 25.04+ host (Kernel 6.12+)
• AMD Ryzen AI Max (Strix Halo)
• Docker with compose plugin

Quick Start

1. ROCm Container

cd rocm
docker compose up -d --build
docker logs rocm-strix

No configuration needed, everything is hardcoded in the compose file.

Verify GPU:

docker exec rocm-strix python3 -c "import torch; print(torch.cuda.get_device_name(0)); print(f'VRAM: {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f} GB')"

2. Ollama Service

cd ollama
cp .env.template .env
# Edit .env, set OLLAMA_MODELS_DIR to a host directory for model storage
mkdir -p /path/to/your/models/dir
docker compose up -d --build
docker logs ollama-strix

OLLAMA_MODELS_DIR points to a directory on your host machine. Models are stored there so they survive container rebuilds. Multiple ollama instances can share the same directory.

Test inference:

curl http://localhost:11434/api/generate -d '{
  "model": "gpt-oss:20b",
  "prompt": "Hello!",
  "stream": false
}'

Ollama Configuration

All configurable via ollama/.env:

Variable	Default	Description
OLLAMA_MODELS_DIR	(required)	Host directory for shared model storage
OLLAMA_MODEL	gpt-oss:20b	Model to auto-download on first start
OLLAMA_CONTEXT_LENGTH	8192	Context window size in tokens
OLLAMA_KEEP_ALIVE	5m	How long models stay loaded in VRAM (-1 = forever)

---

License

MIT for the build glue in this repository (Dockerfiles, docker-compose.yml files, entrypoint scripts, and configuration). Third-party components fetched at build or runtime (PyTorch ROCm wheels, Ollama, model weights pulled by Ollama) retain their own upstream licenses; this repository does not redistribute them.