Fuzzy provides numerical libraries compiled with Verificarlo and its PRISM backend. Run a calculation repeatedly with stochastic rounding to measure its sensitivity to rounding errors, or select a plain build for comparison.
The v2.7.0 recipes build one image per x86-64 CPU level. Each contains ieee
and prism builds of LAPACK, CPython, NumPy, SciPy, scikit-learn, PyTorch,
torchaudio, torchvision and libm. Select each package independently at run time.
| CPU level | Image tag | Alias |
|---|---|---|
| x86-64 baseline (SSE2) | v2.7.0-x86-64-v1 |
v2.7.0-sse2 |
| x86-64-v2 (SSE4.2) | v2.7.0-x86-64-v2 |
v2.7.0-sse4 |
| x86-64-v3 (AVX2) | v2.7.0-x86-64-v3 |
v2.7.0-avx2 |
| x86-64-v4 (AVX-512) | v2.7.0-x86-64-v4 |
v2.7.0-avx512 |
Choose a level your CPU supports. Use v1 when portability matters; latest
points to v1 after all four release images pass validation. Each release is
also mirrored to the GitHub Container Registry with identical digests, e.g.
ghcr.io/verificarlo/fuzzy:v2.7.0-x86-64-v3.
docker run --rm -it verificarlo/fuzzy:v2.7.0-x86-64-v3
fuzzy list
fuzzy use lapack=ieee numpy=prism
python3 script.pyThe per-stack tags (-lapack, -lapack-python…, -numpy, -scipy,
-sklearn, -pytorch…) are legacy. Their recipes and tags remain available;
see the legacy guide. New releases publish the CPU-level tags.
Every package defaults to prism, with stochastic rounding enabled.
fuzzy use python=ieee libm=ieee lapack=ieee numpy=prism
fuzzy run lapack=prism -- python3 script.py
eval "$(fuzzy env python=prism)"fuzzy use changes native library symlinks and installs the chosen Python
wheels with uv. fuzzy run and fuzzy env select native libraries and the
interpreter without writing to the image. On read-only images, including
Apptainer, create a writable Python environment to select Python wheels:
uv venv ~/fuzzy-env --python /opt/fuzzy/python/3.12/bin/python3
fuzzy use --python ~/fuzzy-env numpy=ieee scipy=prism
fuzzy run python=ieee lapack=ieee libm=ieee -- ~/fuzzy-env/bin/python script.pyMount your own code with docker run -v "$PWD:/workdir" …, or derive an
image with FROM verificarlo/fuzzy:v2.7.0-x86-64-v3.
CPython 3.12 uses compensated summation for sum() of floats, which can
recover rounding errors. Use an explicit loop to see variation:
python3 -c '
for run in range(3):
s = 0.0
for _ in range(1000):
s += 0.001
print(s.hex())'PRISM changes inexact operations; exact operations stay exact. Run your real calculation several times and compare its outputs. Fix the seed to reproduce one rounding sequence, select round-to-nearest, or reduce virtual precision:
VFC_BACKENDS="libinterflop_prism.so --mode=sr --seed=42" python3 script.py
VFC_BACKENDS="libinterflop_prism.so --mode=rn" python3 script.py
VFC_BACKENDS="libinterflop_prism.so --precision-binary64=40" python3 script.pyPRISM's round-to-nearest can differ from the plain build in the last bits. MCA modes such as precision bounding are documented in the legacy guide; the v2.7.0 package recipes use PRISM stochastic rounding.
docker/fuzzy/build.sh v3 podman # or docker; v1, v2 and v4 also supportedBuilding runs the compiled code, so v4 requires an AVX-512 builder. CI uses
[self-hosted, Linux, X64, avx512] for v4 and hosted runners for v1–v3.
Each build tests both flavors, mixed selections, runtime sharing and linking,
round-to-nearest agreement, stochastic variation and ISA limits. CI also runs
v1–v3 under Intel SDE with older CPU profiles before publishing.
See the image guide for package versions, Python wheel indexes, supported CPU backends and build details. A hosted wheel index is optional; all wheels ship inside each image.
For instructions on how to contribute, please refer to the Contribution Guide.
If you prefer a visual illustration of everything we've written above, feel free to check-out the slides we recently presented at Scipy2021!
The Fuzzy ecosystem has emerged from — and been used in — several scientific publications. Below is a list of papers which present the techniques used, the tools which have been developed accordingly, and demonstrate how decision-making and applications can be built atop them:
Parker, Douglas Stott, Brad Pierce, and Paul R. Eggert. "Monte Carlo arithmetic: how to gamble with floating point and win." Computing in Science & Engineering 2.4 (2000): 58-68.
Frechtling, Michael, and Philip HW Leong. "Mcalib: Measuring sensitivity to rounding error with monte carlo programming." ACM Transactions on Programming Languages and Systems (TOPLAS) 37.2 (2015): 1-25.
C. Denis, P. De Oliveira Castro and E. Petit, "Verificarlo: Checking Floating Point Accuracy through Monte Carlo Arithmetic," in 2016 IEEE 23nd Symposium on Computer Arithmetic (ARITH), Silicon Valley, CA, USA, 2016 pp. 55-62.
Chatelain, Yohan, et al. "VeriTracer: Context-enriched tracer for floating-point arithmetic analysis." 2018 IEEE 25th Symposium on Computer Arithmetic (ARITH). IEEE, 2018.
Kiar, Gregory, et al. "Comparing perturbation models for evaluating stability of neuroimaging pipelines." The International Journal of High Performance Computing Applications 34.5 (2020): 491-501.
Sohier, Devan, et al. "Confidence Intervals for Stochastic Arithmetic." ACM Transactions on Mathematical Software (TOMS) 47.2 (2021): 1-33.
Kiar, Gregory, et al. "Data Augmentation Through Monte Carlo Arithmetic Leads to More Generalizable Classification in Connectomics." bioRxiv (2020).
Chatelain, Yohan, et al. "PyTracer: Automatically profiling numerical instabilities in Python." IEEE Transactions on Computers (2022).
Des Ligneris, Morgane, et al. "Reproducibility of tumor segmentation outcomes with a deep learning model." 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI). IEEE, 2023.
Pepe, Inés Gonzalez, et al. "Numerical Uncertainty of Convolutional Neural Networks Inference for Structural Brain MRI Analysis." International Workshop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging. Cham: Springer Nature Switzerland, 2023.
Chatelain, Yohan, et al. "A numerical variability approach to results stability tests and its application to neuroimaging." arXiv preprint arXiv:2307.01373 (2023).
Vila, Gaël, et al. "The Impact of Hardware Variability on Applications Packaged with Docker and Guix: a Case Study in Neuroimaging." (2024).
Gonzalez Pepe, Inés, et al. "Numerical stability of DeepGOPlus inference." Plos one 19.1 (2024): e0296725.
The Fuzzy copyright belongs to all contributors of this repository, and it is licensed for public use under the same terms as the LLVM project, which is a modified version of the Apache 2.0 license.