# Installation and compatibility

<a id="rust"></a>

## Rust

Shrinkage's core library requires Rust 1.84 or newer. Until the first crates.io release, use a checkout of the repository. The default build enables no matrix backends and has no Python dependencies. Select the versioned feature matching your matrix library. With the checked-in lockfile, the integrations require: 

| Feature | Integration | Minimum Rust |
| --- | --- | --- |
| `faer_v0_24` | faer 0.24 | 1.84 |
| `nalgebra_v0_34` | nalgebra 0.34 and nalgebra-sparse 0.11 | 1.87 |
| `ndarray_v0_17` | ndarray 0.17 | 1.84 |
| `sprs_v0_11` | sprs 0.11 CSC through `SprsCsc` | 1.88 |

The short names `faer`, `nalgebra`, `ndarray`, and `sprs` are aliases. Optional `experimental-block-reader` and `zarrs_v0_22` features enable the provisional buffered file-backed consumer, which requires Rust 1.89 with the checked-in lockfile. The experimental block reader alone supports Rust 1.84. The documentation reference includes these optional capabilities. 

From a checkout, run the existing fitting example: 

```console
devenv shell -- cargo run --locked --example lasso --features faer_v0_24
```

<a id="python"></a>

## Python

The distribution and import names are both `shrinkage`. The package currently fits dense, single-response Gaussian and binomial models with clustered lasso, chain fused lasso, or SLOPE. Compose an explicit hybrid or proximal-gradient solver, or use the Gaussian lasso convenience API. Results provide prediction, fitted preprocessing, and convergence diagnostics. NumPy 2.0 or newer is required. 

The CI matrix builds an interpreter-specific wheel and tests its installation and import for each supported combination: 

| Interpreter | Linux | macOS | Windows |
| --- | --- | --- | --- |
| Standard CPython 3.11–3.15 | x86\_64 | arm64 and x86\_64 | x86\_64 |

Other interpreters, free-threaded CPython, Linux musl, and additional architectures are outside this initial matrix. These are source-build and CI targets; portable registry wheels are a later distribution milestone. CI runs fitting and prediction tests on installed wheels, including parity with Rust fixtures. 

Build and install the checkout with Rust 1.85 or newer and a supported Python: 

```console
devenv shell
task python:install
uv run --locked --project packages/python python -c "import shrinkage; print(shrinkage.__version__)"
```

The build uses the Rust sources from this repository. Installing an extracted source distribution also works without the repository because Maturin includes the required local Rust dependency. Neither package is published to a registry yet. 

Dense input must consist of exact NumPy arrays with native-endian `float64` dtype, aligned storage, and finite values. `X` has shape `(n_samples, n_features)`, and `y` has shape `(n_samples,)`, with at least one observation. Preparation copies C, Fortran, and strided layouts into independent Rust-owned storage. Read-only arrays are copied too, and requests for zero-copy NumPy input are rejected. See the [full Python input contract](<https://github.com/jolars/shrinkage/blob/main/packages/python/README.md#dense-input-contract>) for ownership, copying, and validation details.