Installation¶

System Requirements¶

Component	Why it is needed
Docker + docker compose	Runs the `app` and `ollama` services defined in `compose.yml`.
NVIDIA Container Toolkit	Exposes host GPUs to the `ollama` container.
At least 40 GB free disk	LLM weights (e.g. ~40 GB for the 70 B Llama).
`obolibrary/robot:latest` Docker image	Required by Ontology Preparation (auto-pulled on first use).
`gawk` on the host	Used by the Plant Ontology preprocess step in `scripts/build_subset_ontologies.sh`. The ROBOT image ships `mawk`, which lacks `gensub()`. Install with `apt-get install gawk` (Debian/Ubuntu) or `brew install gawk` (macOS).
Python 3.10+ and uv	Only for local (non-Docker) development.

Docker Compose (Recommended)¶

Setup¶

git clone https://github.com/dbcls/bsllmner-mk2.git
cd bsllmner-mk2
docker compose up -d --build

This builds the app image, pulls ollama/ollama:0.32.15, and starts both containers on the bsllmner-mk2-network bridge. Named volumes:

bsllmner-mk2_venv -- the project's .venv inside the app container.
bsllmner-mk2_ollama-data -- Ollama's model cache (/root/.ollama).

The host's project directory is bind-mounted at /app, so source edits take effect without rebuilding.

GPU Configuration¶

compose.yml reserves every visible NVIDIA GPU (deploy.resources.reservations.devices.count: all) and the ollama service narrows the active set with CUDA_VISIBLE_DEVICES. The committed default is 0,1 (two GPUs). On a single-GPU host:

# compose.yml (ollama service)
environment:
  - CUDA_VISIBLE_DEVICES=0

For Ollama performance tuning (KV cache, flash attention, context length tiers), see Configuration.

uv (Local Development)¶

For running tests, type checks, or the CLI directly on the host:

uv sync --all-extras

If the Ollama server is on a different host, set OLLAMA_HOST:

export OLLAMA_HOST=http://<ollama-host>:11434

Verify Installation¶

# Docker Compose
docker compose ps
docker compose exec app bsllmner2_extract --help

# uv
uv run bsllmner2_extract --help