pytest-numguard

pytest plugin for numerical edge-case testing of array code: it runs your functions on NaN, ±inf, huge and subnormal magnitudes, cancellation-prone data, empty arrays and float16/32/64 inputs, and compares them against references and metamorphic properties.

pip install pytest-numguard

This package installs numguard-ci, which provides the plugin (the command and the import are numguard). Once it is installed, pytest picks the plugin up automatically. The example needs SciPy for the reference (pip install scipy):

import numpy as np
import pytest
from scipy.special import logsumexp as reference


def naive_logsumexp(x):
    return np.log(np.sum(np.exp(x)))


def stable_logsumexp(x):
    x = np.asarray(x)
    if x.size == 0:
        return x.dtype.type(-np.inf)
    m = np.max(x)
    if not np.isfinite(m):
        return m
    return (m + np.log(np.sum(np.exp(x - m)))).astype(x.dtype)


@pytest.mark.numguard(fn=naive_logsumexp, reference=reference)
def test_naive_logsumexp():
    pass


@pytest.mark.numguard(fn=stable_logsumexp, reference=reference, properties=["translation_equivariant"])
def test_stable_logsumexp():
    pass

Run pytest: the naive version fails on the huge, large, cancellation and wide_range cases (it returns inf where the reference is finite), and the stable version passes. Options such as --numguard-backend torch and the numguard.toml configuration are documented at https://github.com/bhaskargurram-ai/numguard.