Installation - SiNDAE

SiNDAE is published on PyPI and installs with pip. conda is offered as an alternative for platforms where the optional binary dependencies (cyipopt, mpi4py) are easier to obtain as pre-built packages.

Requirements¶

Python 3.11 or newer
pip (recommended) or conda
Linux, macOS, or Windows (WSL2 recommended on Windows)

We recommend installing into a fresh virtual environment so SiNDAE and its dependencies do not interfere with other projects:

python -m venv .venv
source .venv/bin/activate       # macOS / Linux
# .venv\Scripts\activate        # Windows

Install with pip (recommended)¶

Core install¶

pip install sindae

This pulls the core stack (numpy, scipy, jax, equinox, optax, pyomo, matplotlib) together with the pure-Rust solvers POUNCE and FERAL, which install from wheels with no system libraries and no license. The core install runs the entire SiNDAE pipeline (problem, smoother, simultaneous and decomposition training, grey-box models, inference, and plotting) on POUNCE and FERAL.

Full install¶

pip install "sindae[full]"

The full extra adds mpi4py, required for MPI-parallel decomposition training, and cyipopt, an optional alternative NLP backend (POUNCE is the default for every solve). Their wheels depend on your platform; if either fails to build from pip, use the conda route below for those two packages.

Model export (optional)¶

pip install "sindae[onnx]"    # HybridDAE.export to ONNX (+ a scaler sidecar) or JSON
pip install "sindae[omlt]"    # HybridDAE.to_omlt for an in-memory OMLT model

These extras are only needed to hand a trained network to a foreign optimization tool. The core pipeline never imports them, and HybridDAE.export(..., format="json") needs no extra at all. See the HybridDAE API page for usage.

Install with conda (alternative)¶

conda-forge ships pre-built binaries for the dependencies that are awkward to build with pip: cyipopt (bundled with IPOPT and MUMPS, no HSL license) and mpi4py (linked against OpenMPI). Install those with conda, then install SiNDAE itself with pip:

conda create -n sindae python=3.11
conda activate sindae
conda install -c conda-forge cyipopt mpi4py
pip install sindae

Packages¶

Component	Install	Purpose
`numpy`, `scipy`, `matplotlib`	core	Numerics and plotting
`jax`, `jaxlib`	core	Automatic differentiation (CPU by default)
`equinox`, `optax`	core	Neural network layers and optimizers
`pyomo`	core	Symbolic DAE model building and collocation
`pounce-solver`	core	IPOPT-compatible NLP solver (pure-Rust wheels, no HSL)
`feral-solver`	core	Sparse symmetric KKT solver for the decomposition gradient
`cyipopt`	`[full]`	Optional alternative NLP backend (IPOPT with MUMPS); POUNCE is the default
`mpi4py`	`[full]`	MPI parallelism for multi-trajectory decomposition training
`jax2onnx`, `onnxruntime`	`[onnx]`	Export a trained network to ONNX (`HybridDAE.export`)
`omlt`	`[omlt]`	Export a trained network as an OMLT model (`HybridDAE.to_omlt`)

GPU and accelerator support (optional)¶

By default JAX runs on CPU. To use a GPU or Apple Silicon, install the matching JAX build after installing SiNDAE:

pip install -U "jax[cuda12]"     # NVIDIA GPU (CUDA 12)
pip install -U "jax[metal]"      # Apple Silicon (Metal)

Development install (from source)¶

To contribute or to run the test suite, install an editable copy from a clone. The test extra adds pytest:

git clone https://github.com/llueg/SiNDAE.git
cd SiNDAE
pip install -e ".[full,test]"

The repository also ships an environment.yml for a one-command conda development environment (Python, cyipopt, mpi4py, and an editable install):

conda env create -f environment.yml
conda activate sindae

Run the fast test suite with:

pytest

Verify the installation¶

import sindae
import jax
import pyomo.environ as pyo

jax.config.update("jax_enable_x64", True)
print("JAX devices:", jax.devices())

# POUNCE backs the simultaneous workflow (core install).
print("POUNCE available:", pyo.SolverFactory("pounce").available())

# cyipopt is optional: a selectable alternative NLP backend, not required for any feature.
try:
    print("cyipopt available:", pyo.SolverFactory("cyipopt").available())
except Exception as exc:
    print("cyipopt not available:", exc)

Troubleshooting¶

pounce not available. Make sure your virtual environment is active. pounce-solver installs the pounce executable into the environment’s bin directory, which must be on PATH.

cyipopt fails to build from pip. Its wheels are platform-specific. Install it from conda-forge instead (conda install -c conda-forge cyipopt), then pip install sindae into the same environment.

JAX emits float32 warnings. Set 64-bit mode at the top of your script with jax.config.update("jax_enable_x64", True), or set JAX_ENABLE_X64=1 in your shell.

mpirun: command not found. Install an MPI implementation: OpenMPI via conda install -c conda-forge openmpi, brew install open-mpi (macOS), or sudo apt install libopenmpi-dev (Ubuntu).