Installation

Prerequisites

Requirement	Why
Python 3.12+	PEP-695 generics used internally
uv ≥ 0.5	The only supported dependency manager
NVIDIA GPU + CUDA 12.x	Required for training / evaluation
Git	to clone the repository

Verify your GPU once with nvidia-smi. If it fails, training will crash at the first .cuda() call.

A. uv (recommended)

git clone https://github.com/jajupmochi/swiss-river-network-benchmark.git
cd swiss-river-network-benchmark
uv sync --no-cache
uv run srn --help

Optional extras:

uv sync --all-extras              # everything
uv pip install -e '.[app]'        # demo apps only
uv pip install -e '.[docs]'       # mkdocs + i18n + mike
uv pip install -e '.[dev]'        # ruff + pytest + nbmake

B. pip

python -m pip install 'swissrivernetwork[app]'        # once on PyPI
pip install -e '.[app]'                               # from a clone

C. Docker

docker compose --profile ui up -d app     # Streamlit on :8501
docker compose --profile train run --rm train srn sweep

See docker-compose.yml for all profiles (train, ui, notebook).

Sanity check

uv run srn version
uv run python -c "import torch; assert torch.cuda.is_available()"

Both lines should exit cleanly.