Dakota wheel built by the IT'IS Foundation

This repository contains the code to build a Python wheel to load the Dakota python module.

Installing the wheel

pip install itis-dakota

Requirements

Pre-built wheels are published on PyPI for these platforms, for CPython 3.12–3.14:

Platform	Wheel tags	Notes
Linux	manylinux_2_28 (x86_64, aarch64)	manylinux compatible distributions
macOS	macosx_*_arm64, macosx_*_x86_64
Windows	win_amd64	built in CI with an MSYS2/MinGW-w64 (ucrt64) toolchain; runs on any 64-bit Windows with the Universal CRT (Windows 10/11 out of the box). Not yet on PyPI: these wheels (~80 MB) exceed PyPI's default 60 MB per-file limit, so pip install availability starts after a one-time project limit increase and the next release

On Windows, install a regular 64-bit CPython from python.org (or any virtual environment based on it, e.g. venv or conda); the dakota.exe launcher and its bundled runtime DLLs are installed next to the interpreter and need no separate compiler or Visual C++ redistributable. The Windows "embeddable zip" distribution is untested; use the full installer.

Known limitation: Dakota's binary restart files are not portable across platforms (Boost archives are native-format checked, e.g. sizeof(long) differs between Linux and Windows).

Using the wheel

After installation, the module can be imported:

import dakota
import dakota.environment as dakenv

And example on how to use the environment module can be found here: https://github.com/snl-dakota/dakota/blob/devel/src/unit/test_dakota_python_env.py

Building the wheel

make wheel

Typing / stubs

dakota.environment is a compiled pybind11 extension, so its type stubs (stubs/dakota/environment/{__init__.pyi,environment.pyi}) are generated, not hand-written. After bumping the Dakota version or touching dakota/src/dakota_python.cpp via a patch, regenerate the stubs, then commit the result:

make stubs

make stubs builds the wheel first when wheelhouse/ is empty. Commit the regenerated stubs: the wheel used for generation still ships the previous ones, and wheels built afterwards bundle the committed stubs.

CI fails (one linux matrix leg) if the committed stubs no longer match the compiled module.

Dakota's native JSON input format

Since Dakota 6.24, studies can be configured directly from JSON instead of Dakota's legacy keyword input file. Dakota's own Pydantic v2 models for that JSON format (dakota/python/dakota/spec, source of truth for dakota.json and the C++ JSON parser) are shipped as dakota.spec, giving IDE autocomplete/inline docs and real validation without a round-trip to the online docs:

from dakota.spec.study import DakotaStudy
import dakota.environment as dakenv

study = DakotaStudy(
    method=[...],
    variables=[...],
    responses=[...],
)
dakenv.study(callback, study.model_dump(mode="json", exclude_none=True))

dakota.spec is pure Python (no compiled dependency) and is copied verbatim from the vendored Dakota source on each build, so it stays in sync with whatever Dakota version this wheel bundles automatically.

Copyright (c) 2023-2026 IT'IS Foundation, Zurich, Switzerland