fcmaes-rust for Python

Choosing an optimization tool? Start with the shared solver-selection and agent guide or the optimization ecosystem overview. This distribution exposes the Rust optimizers to Python. For parallel Python callbacks, compare with fcmaes (fast-cma-es); the job worker and experimental surrogate are source-only components, not part of this wheel.

fcmaes-rust provides Python access to fast, parallel, gradient-free optimization algorithms implemented in native Rust. The optimizer core is pure Rust and does not compile, link, or load the historical C++ backend.

Installation

python -m pip install fcmaes-rust

The initial release supports CPython 3.11 through 3.13 on the platforms for which wheels are listed on PyPI.

To build from source, install Rust 1.88 or newer (https://rustup.rs), then either let pip build the source distribution (python -m pip install --no-binary fcmaes-rust fcmaes-rust) or build a checkout of the repository:

git clone https://github.com/dietmarwo/fcmaes-rust
cd fcmaes-rust
python -m pip install .

The Python package depends on NumPy and SciPy and, at build time, on crates.io crates only.

Quick check

import fcmaes_rust

print(fcmaes_rust.__version__)
print(fcmaes_rust.phase1_build_info())

Optimizers

The package exposes one-shot and ask/tell interfaces for Differential Evolution, active CMA-ES, CR-FM-NES, PGPE and BiteOpt, plus Dual Annealing, MODE, MAP-Elites, the Diversifier, independent retry, coordinated retry and weighted multi-objective retry.

The functions currently return low-level NumPy arrays, tuples or dictionaries. See the Python binding guide for signatures, examples, parallelism behavior and result layouts. Every exported function, stateful method, and property also provides an installed runtime docstring, for example help(fcmaes_rust.optimize_de) or help(fcmaes_rust.MODE). SciPy is a runtime dependency because retry callbacks receive a public scipy.optimize.Bounds object.

import numpy as np
import fcmaes_rust

def sphere(x):
    return float(np.dot(x, x))

dim = 4
empty = np.empty(0, dtype=np.float64)
x, value, evaluations, iterations, stop = fcmaes_rust.optimize_de(
    sphere,
    dim,
    np.full(dim, -5.0),
    np.full(dim, 5.0),
    empty,
    empty,
    np.empty(0, dtype=np.bool_),
    seed=7,
    max_evaluations=4_000,
    popsize=20,
)
print(value, evaluations, x)

Python callbacks must reacquire the GIL. Native Rust objectives offer the strongest parallel scaling; cheap pure-Python callbacks are usually callback-bound.

Project links

• Source repository
• Documentation
• Issue tracker
• Release history

License

MIT