Installation

Install from PyPI

The recommended way to install pySigLib is from PyPI using pre-built wheels. No compiler toolchain is required - wheels are published for Windows, Linux, and macOS (arm64).

pip install pysiglib              # CPU only
pip install pysiglib[cuda]        # with CUDA GPU support

The [cuda] extra installs the companion pysiglib-cuda plugin, which ships the CUDA shared library (cusig) and the CUDA JAX FFI bindings as a sibling package. pysiglib discovers it at import time; if the plugin is absent, pysiglib runs CPU-only.

JAX support

The XLA FFI bindings for JAX are already built into every pySigLib wheel, so there is no pySigLib-side extra to install. Just install JAX separately:

pip install jax

For GPU-accelerated JAX, use its CUDA variant:

pip install pysiglib[cuda] jax[cuda12]

To verify the installation:

import pysiglib
print(pysiglib.__version__)
print(pysiglib.BUILT_WITH_CUDA)    # True if CUDA backend loaded
print(pysiglib.BUILT_WITH_JAX_FFI) # True if JAX FFI available

Install from source

If you need a custom build (unsupported platform, alternative CUDA version, development work), pySigLib can be built from source. pysiglib and the pysiglib-cuda plugin are separate packages with separate builds: install pysiglib for the CPU core, and additionally install pysiglib-cuda from the plugins/cuda subdirectory of a repo checkout for the CUDA backend.

Base package (CPU only)

This requires a C++ compiler toolchain. The base wheel is CPU-only - it does not build CUDA. Use the pysiglib-cuda plugin for that (next section).

Windows

Requires MSVC. Once installed, run:

pip install pysiglib --no-binary pysiglib

Linux

Requires GCC >= 10. Once installed, run:

pip install pysiglib --no-binary pysiglib

macOS

Requires the Xcode Command Line Tools (xcode-select --install). Once installed, run:

pip install pysiglib --no-binary pysiglib

CUDA plugin

The pysiglib-cuda plugin is published as wheels only - there is no sdist on PyPI, because it depends on sources from the parent repository (siglib/cusig). To build the plugin from source, clone the repository and install the plugins/cuda subdirectory:

git clone https://github.com/daniil-shmelev/pySigLib.git
cd pySigLib
pip install pysiglib --no-binary pysiglib
pip install ./plugins/cuda

This requires a working CUDA toolkit (nvcc on PATH or CUDAToolkit_ROOT set to its install prefix; on Linux this is typically /usr/local/cuda, and on Windows the CUDA_PATH environment variable set by the NVIDIA installer is also picked up).

By default the plugin compiles only for the local GPU’s architecture (CUDA_ARCH=native). To target multiple architectures, set CUDA_ARCH to a semicolon-separated list, all-major, or all before installing:

CUDA_ARCH="80;89;90" pip install ./plugins/cuda

JAX support (source builds)

When building from source, both packages automatically detect JAX and build their respective XLA FFI bindings if JAX is installed. Requires jaxlib >= 0.9.1 (Python 3.11+):

pip install jax
pip install pysiglib --no-binary pysiglib
pip install ./plugins/cuda    # only if you need CUDA

If JAX is not installed at build time, the FFI bindings are skipped and the rest of pySigLib works normally. To verify:

import pysiglib
print(pysiglib.BUILT_WITH_JAX_FFI)

Build options

The following environment variables control the source build:

Base pysiglib build

Variable	Default	Description
PYSIGLIB_JAX_FFI	ON	Set to 0 to disable JAX FFI support. When ON (default), JAX FFI is built automatically if JAX is installed, and skipped otherwise.
SIGLIB_VEC	ON	Set to 0 to disable AVX vectorization.

pysiglib-cuda plugin build

Variable	Default	Description
CUDA_ARCH	native	CUDA architectures to compile for. Accepts native (local GPU only), all (all architectures), all-major, or a semicolon-separated list (e.g. "80;89;90"). Use all when building portable wheels.
CUDAToolkit_ROOT	(auto)	CUDA toolkit prefix. On Windows CUDA_PATH (set by the NVIDIA installer) is picked up automatically; set this only if CMake cannot locate nvcc on its own.

Editable installs

Both packages support editable installs for development:

pip install -e .                  # base pysiglib
pip install -e ./plugins/cuda     # CUDA plugin

Citation

If you found this library useful in your research, please consider citing the paper:

@article{shmelev2025pysiglib,
  title={pySigLib-Fast Signature-Based Computations on CPU and GPU},
  author={Shmelev, Daniil and Salvi, Cristopher},
  journal={arXiv preprint arXiv:2509.10613},
  year={2025}
}