Conda environments for Exudyn

Notes on creating and maintaining conda environments for Exudyn — for development, for documentation generation, and for running the examples.

• Author: Johannes Gerstmayr

• Date: 2026-09-09

For basic setup of Anaconda or Miniconda itself, see the Exudyn documentation.

Dependency policy

numpy is the only hard requirement — it follows from the nature of the C++/Python coupling. Everything else is optional: installing Exudyn pulls the minimum, and functions needing more raise at the point of use. The lists below are therefore conveniences for running examples, not install requirements.

package	needed for
numpy	required, always
scipy	FEM, eigenvalue computations, sparse solvers — pin to 1.15.2, see note below
matplotlib	plotting in examples and exudyn.plot
ngsolve, h5py	FEM mesh import and FEMinterface
numpy-stl	STL import/export
numba	a few accelerated example kernels
torch, stable-baselines3	reinforcement learning examples
mpi4py	parallel parameter-variation examples
tqdm, psutil	progress bars and process info in exudyn.processing
ipywidgets, ipykernel	Jupyter notebook usage
spyder-kernels	Spyder integration (version must match Spyder, see below)

The standard environment

venvExuP313 is the reference environment: it runs Exudyn, the generators, the tests and the documentation build. Its packages are not listed here but in pyproject.toml, as dependency groups (PEP 735), so this recipe cannot fall behind:

conda create -n venvExuP313 python=3.13 -y
conda activate venvExuP313
cd EXUDYN_git                                 #the repository root, where pyproject.toml is
python -m pip install --upgrade pip           #dependency groups need pip >= 25.1
pip install --group dev                       #docs, lint, build, test and IDE tools, scipy pin, jinja2, griffe
pip wheel . -w dist --no-deps                 #build Exudyn
pip install --pre --find-links=dist "exudyn[tests]"   #the local wheel plus what TestModels/ needs

group	contains	used by
docs	sphinx and its theme and extensions	sphinx-build, the CI docs jobs, readthedocs
pdf	nothing from pip; the flow charts are committed as PDF and SVG	exudev docs --pdf, which needs a LaTeX installation; exudev figures, which compiles a changed flow chart, too
lint	pydoclint (pinned; the baseline holds its messages)	CI check_docstrings
build	setuptools>=77, wheel, pybind11<3.0, tomli, cibuildwheel	a direct python setup.py bdist_wheel, which has no build isolation; a local CI-identical wheel
test	pytest, pytest-xdist, quickjs	pytest -n 8; the suite itself runs without them
ide	spyder-kernels, ipykernel, ipywidgets	Spyder and Jupyter
dev	all of the above, plus scipy==1.15.2 (any scipy on macOS and Python 3.14+), jinja2, griffe	the development environment

Groups are never published with the wheel, unlike the extras (exudyn[tests] etc.) below, which users see. A single group can be installed alone, e.g. pip install --group docs.

Why pybind11 in build? It is declared in build-system.requires, so pip wheel . and python -m build install it themselves under build isolation. A direct python setup.py bdist_wheel does not use build isolation, so there it has to be present in the environment.

Why scipy 1.15.2? scipy 1.18.0 slows the Exudyn test suite from ~22 s to over 10 minutes, apparently in the eigensolver path. The pin is in the dev group; exudyn[tests] itself does not pin scipy. It does not apply on macOS and on Python 3.14 and later, where scipy 1.15.2 has no wheel and pip install --group dev would fail (#2886); there the suite may be slower.

Linux: jagged fonts in the dialogs? The tk package of the Anaconda channel is built without Xft, so tkinter draws its fonts without anti-aliasing and in few sizes; the dialogs of Exudyn say so once when they open. The Xft build of conda-forge draws them smoothly:

conda install -c conda-forge "tk=*=xft_*"

A system Python (apt install python3-tk) has Xft already (#2890).

spyder-kernels in the ide group is 3.*, matching Spyder 6; for an older Spyder see the table below and install the matching version afterwards.

Test environments per Python version are named venvP310 … venvP314 and carry an Exudyn build for that version. --py all of exudev means these five; a newer Python is used when named, e.g. --py 315 with venvP315 (3.15 and 3.16 are accepted), until wheels are built for it (#2891).

Documentation toolchain only

If an existing environment only needs the docs tools added, from the repository root:

pip install --group docs

Build the HTML documentation from the repository root:

sphinx-build -b html . _build -E

Optional packages via the Exudyn extras

The optional dependencies are declared in pyproject.toml, so they can be installed by name instead of being listed by hand:

command	installs
pip install exudyn[basic]	what most models need: scipy (eigenvalues, sparse matrices, FEM) and matplotlib (plots, PlotSensor) - recommended to start with
pip install exudyn[tests]	the above plus exactly what TestModels/ needs: h5py, networkx, psutil, ngsolve
pip install exudyn[common]	the above plus what frequently used features need: numpy-stl (STL import), tqdm (optimization progress), ffmpeg-python (video export; the program itself: conda install ffmpeg) - to use most features
pip install exudyn[all]	the above plus everything else the package and the Examples refer to: pymeshlab, roboticstoolbox-python, spatialmath-python, numba, dispy, mpi4py
pip install exudyn[rl]	reinforcement learning: torch, stable-baselines3, gymnasium, gym, tensorboard

[rl] is deliberately not part of [all]: torch is multi-GB and the CPU/CUDA choice is made with an --index-url, which cannot be expressed in wheel metadata. For a CUDA build, install it separately first:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install stable-baselines3[extra]

The --index-url above selects the CUDA 11.8 build of PyTorch; drop it for a CPU-only install.

These lists are not maintained by hand. python tools/checkExtras.py scans every import in exudyn/, TestModels/ and Examples/ and fails if one of them is installed by no extra, so the table above cannot silently fall behind the code; CI runs it as the check_extras job.

Spyder kernel versions

spyder-kernels must match the Spyder version, or the console fails to start:

Anaconda / Spyder	spyder-kernels
Anaconda 2023-07 / Spyder 5.4	spyder-kernels=2.4
Anaconda 2023-09 / Spyder 5.5, 5.5.1	spyder-kernels=2.5
Anaconda 2025-06 / Spyder 6.0.7	spyder-kernels=3.0

Useful conda commands

conda activate ENVNAME              # activate an environment
conda deactivate                    # leave it

conda env list                      # list all environments
conda list                          # list packages in the active environment

conda install numpy                 # install into the active environment
conda install -n ENVNAME numpy      # install into a named environment
conda update numpy                  # update in the active environment
conda remove --name ENVNAME numpy   # remove a package

conda remove --name ENVNAME --all -y   # delete an entire environment
conda update -n base -c defaults conda -y   # update conda itself

manylinux wheels (auditwheel)

Now normally done with Docker; kept here for reference.

pip install auditwheel patchelf

auditwheel repair /mnt/c/DATA/cpp/Exudyn_git/dist/exudyn-1.10.19.dev1-cp314-cp314-linux_x86_64.whl
# produces e.g. exudyn-1.10.19.dev1-cp314-cp314-manylinux_2_34_x86_64.whl