Installation

From PyPI

This is the usual method. It gives you the milkwashy command and the milkwashr extension module.

uv pip install milkwashy

pip install milkwashy also works. The wheel for milkwashr needs no compiler, because one wheel serves all CPython versions from 3.9.

Make sure that the installation works:

milkwashy config
python -c "import milkwashr; print(milkwashr.ping())"

The second command must print milkwashr-ok.

Only the chunk code

Install milkwashr alone if you do not need the sweep:

uv pip install milkwashr        # Python
cargo add milkwashr             # Rust

For Rust, switch off the default features. This removes PyO3 from your build:

[dependencies]
milkwashr = { version = "1", default-features = false }

From this repository

The script builds both wheels and installs them into the interpreter that uv finds as the system Python. The milkwashy command is then on your PATH.

git clone https://github.com/twardoch/milkwash
cd milkwash
./install.sh

To use a different interpreter, name it:

./install.sh --python /opt/homebrew/bin/python3.12

The script tests the wheels in a temporary environment before it installs them. It also makes sure that it changes no file in the repository.

For development

uv venv --python 3.12 .venv
source .venv/bin/activate
uv pip install -e ./milkwashr -e ./milkwashy pytest
python -m pytest milkwashy/tests -q

Configuration

Write the configuration file, and then look at it:

milkwashy init_config     # writes the file and prints its location
milkwashy config          # prints the values in use

The file is TOML, and platformdirs decides its location. On macOS the location is ~/Library/Application Support/milkwashy/config.toml. On Linux it is ~/.config/milkwashy/config.toml.

Copy the example environment file if your servers need a key:

cp .env.example .env

LM Studio needs no key. See Model servers for the details.

What you need for a sweep

• A server with an OpenAI-compatible interface. LM Studio is the default.
• Source texts. milkwashy gen_indata writes four example texts.
• Time. A complete sweep over many models takes hours. You can stop it and start it again, because the sweep continues where it stopped.