Installation

Requirements

Python 3.10, 3.12 or 3.13 (these run in CI)
NumPy (installed automatically)
A C compiler such as cc or gcc, only if you want the generated-C kernel backend (Stage 3)
PyTorch, only for sf.torch (Stage 2)

Install

Clone the repository

git clone https://github.com/aloshdenny/superfloat.python
cd superfloat.python

Install in a virtual environment

python3 -m venv .venv && source .venv/bin/activate
pip install -e .

Add [test] to also get pytest and hypothesis, or [torch] for the PyTorch integration:

pip install -e ".[test]"
pip install -e ".[torch]" --extra-index-url https://download.pytorch.org/whl/cpu

Check it works

import sf
sf.sf16(0.5) + sf.sf16(0.25)
# sf16(0.75)

Run the tests

pytest -m "not slow"     # about 40 seconds
pytest                   # the full per-commit suite, about 8 minutes

The quick run skips the exhaustive sf9 and sf10 sweeps. Both runs compare every result against the exact-fraction oracle; see how sf is tested.

Run the tours

Each stage ships a script that walks through its features and prints what it does:

python examples/stage_1_0_tour.py     # scalars and the error policy
python examples/stage_1_1_tour.py     # arrays, matmul, the device path

Everything in this documentation was produced by running the code shown. If an output on these pages differs from what you see, that is a bug in one or the other; please report it.

Get started

Concepts

Scalars

Arrays

Errors

Testing

Requirements

Install

Run the tests

Run the tours

​Requirements

​Install

​Run the tests

​Run the tours

Requirements

Install

Run the tests

Run the tours