Installation¶

Install spectrafit-core (as a user)¶

No wheel is published yet — install from source

spectrafit-core is not on PyPI. No version tag has been cut in this repository either, so uv add spectrafit-core and pip install spectrafit-core will both fail with "no matching distribution". This is the same fact CITATION.cff records by deliberately omitting its date-released field. Until the first release is published, the source build below is the only path that works.

Requires Python \(\geq\) 3.13, plus a Rust toolchain — building from source means compiling the PyO3 extension yourself. Building also compiles a LAPACK dependency: spectrafit-varpro's VarPro solver selects lapack-accelerate on macOS and lapack-netlib (which needs a Fortran compiler and CMake) on every other platform (crates/spectrafit-varpro/Cargo.toml). What that means per platform:

macOS: nothing extra to install — lapack-accelerate links against the OS-provided Accelerate framework.
Linux: a Fortran compiler, CMake, and LAPACK/OpenBLAS headers. This project's own CI image installs build-essential cmake gfortran liblapack-dev libopenblas-dev (.gitlab/docker/Dockerfile.ci); the GitHub release workflow's manylinux wheel build installs liblapack-dev libopenblas-dev on top of the manylinux image's own toolchain (.github/workflows/release.yml).
Windows: unverified by this project's own CI. .github/workflows/ci.yml (the test suite) has no Windows job, so nothing here has actually been tested on Windows. .github/workflows/release.yml does build Windows wheels, but — unlike its Linux step — installs no Fortran/LAPACK toolchain first. Windows wheels are built, not tested; treat Windows support as unverified rather than assumed.

git clone https://github.com/Anselmoo/SpectraFit-Core.git
cd SpectraFit-Core
uv sync
uv run maturin develop --release

--release matters: a debug extension runs 20–50× slower than a release one. Then run your first fit from inside that checkout — uv run python -c "import spectrafit_core" should import cleanly.

Note

Status: beta (0.1.0). APIs and the benchmark contract may still change before the stable 1.0 release — see the project's LIMITATIONS.md for currently disclosed gaps.

Once the first release is published¶

These are the commands to use as soon as a wheel exists on PyPI; they are recorded here so this page needs no rewrite at release time, not because they work today — both currently 404 against PyPI, per the warning above.

Install commands for once a release exists (not live yet)

Using uv (recommended):

uv add spectrafit-core

Or plain pip:

pip install spectrafit-core

At that point no local Rust toolchain will be needed to use the library — the wheel bundles the compiled Rust core.

Either way, continue to Quickstart to run your first fit.

Installing for development¶

If you want to modify spectrafit-core itself (Rust kernel, Python bindings, benchmark engine, or the web dashboard), you'll need the Rust toolchain and maturin to build the extension locally, plus uv to manage the Python environment. Three steps, run once:

flowchart LR
    A["uv sync"] --> B["maturin develop"]
    B --> C["pytest"]

1. Sync Python dependencies¶

uv sync --extra benchmark

Dev tooling (pytest, ruff, ty, the benchmark oracles) is a dependency group installed by default — this one command is enough, no separate --dev flag needed.

2. Build the Rust extension¶

uv run maturin develop

This compiles crates/spectrafit-core (the PyO3 cdylib) and installs it into your uv-managed virtualenv as an editable package. Expect a cold build to take a minute or two; a successful run ends with:

📦 Built wheel for CPython 3.13 to /tmp/.../spectrafit_core-0.1.0-cp313-cp313-macosx_11_0_arm64.whl
✏️ Setting installed package as editable
🛠 Installed spectrafit-core-0.1.0

3. Run the test suite¶

uv run pytest

A healthy run ends with a summary line shaped like this (exact counts drift as the suite grows — the shape is what to expect, not these specific numbers):

==== 1577 passed, 33 skipped, 4 xfailed, 5 xpassed in 6m11s ====

For the full contributor setup — required tool versions, the fast local lint/test loop, and the MCP-first tooling workflow — see the contributor setup guide.

Next steps¶

Run your first fit — Quickstart fits a single Gaussian peak in a few lines against the public API, a quicker check that the build actually works than the test suite alone.