Installation

Several installation methods for MOSFiT are outlined below. If you run into issues, open a new issue on GitHub.

MOSFiT requires Python 3.11 or newer (3.11–3.14 are tested). Runtime dependencies are declared in pyproject.toml (including astrocats>=0.5.0, dynesty>=3.1, NumPy 2, and Astropy 7). There is no setup.py.

Installing from source

Platforms: macOS, Linux, and Windows

For development work, clone the repository and sync the project with uv:

git clone https://github.com/guillochon/MOSFiT.git
cd MOSFiT
uv sync

This creates a local virtual environment in .venv. Run MOSFiT with uv run mosfit ..., or activate the environment and call mosfit directly. A .python-version file pins a modern CPython for local uv use; any interpreter satisfying requires-python = ">=3.11" is allowed.

Optional extras:

uv sync --extra mpi      # mpi4py (needs a system MPI library)
uv sync --extra sedona   # PyTorch, only for the SESN SEDONA emulator
uv sync --extra docs     # Sphinx stack for building these pages

Combine extras as needed, e.g. uv sync --extra mpi --extra sedona.

Equivalent editable pip install from the same pyproject.toml:

pip install -e .
pip install -e ".[mpi,sedona,docs]"

Installing with pip

Platforms: macOS, Linux, and Windows

Installing MOSFiT with pip from PyPI:

pip install mosfit
pip install 'mosfit[sedona]'   # SESN SEDONA emulator
pip install 'mosfit[mpi]'      # mpi4py

Published wheels may lag this development branch.

Setting up MOSFiT with Anaconda

Platforms: macOS, Linux, and Windows

The Anaconda or Miniconda distribution is a convenient Python environment. After installing conda, a released MOSFiT can be installed via:

conda install -c conda-forge mosfit

MOSFiT 2.0 is a hatchling package (no setup.py). After this release is on PyPI, conda-forge needs a feedstock bump: use the template in recipe/ of this repository. astrocats >= 0.5.0 must be on conda-forge first (the channel still has 0.3.37 at the time of writing). PyTorch and mpi4py are optional extras, not required conda run dependencies. Until that feedstock PR lands, install 2.0 from source with uv sync or from PyPI with pip.

Lightweight install for rest-frame SEDs

If all you need is the rest-frame SED path (Rest-frame SEDs for external light-curve simulators) — driving MOSFiT as a source model from an external light-curve simulator, or generating training data — you do not need the sampling stack. Nothing on that path imports emcee, dynesty or numba, and a CI job keeps it that way.

pyproject.toml declares a lynx dependency group holding just that subset. Because extras can only ever add packages, the lightweight install works by suppressing MOSFiT’s own dependencies and then installing the group on its own:

uv venv
uv pip install --no-deps -e .     # or: --no-deps mosfit
uv pip install --group lynx

This drops dynesty, numba and llvmlite — the last two being by far the largest of the omitted wheels. emcee is kept, since the command-line --lynx run draws from the priors with the ensembler; mosfit.lynx. LynxSource does not need it.

Two caveats worth stating plainly:

astrocats imports matplotlib, seaborn and pandas at module scope, so those arrive regardless of what MOSFiT asks for. They are not something this install path can avoid.
The tde model JIT-compiles its viscous transform, so it additionally needs numba. Every other bundled model runs without it.

Anything beyond the SED path — fitting real events, plotting, nested sampling — wants the full install described above.

Docker

A CPU runtime image is built from the repository Dockerfile. It installs MOSFiT with uv sync --frozen --extra mpi on Python 3.14.7 and does not include PyTorch. import mosfit works without the sedona extra. Local fits inside the image can use --max-cores; MPI is available when jobs are launched with mpirun. Build with Docker from the repository root; see .github/workflows/docker-publish.yml for the published image tags.