meltygui-pycuda

Namespaced pycuda bindings maintained for MeltyGUI, based on the upstream source recorded in UPSTREAM.json. Original authorship and licenses are preserved. This repository contains the binding source and its release tooling.

uv pip install meltygui-pycuda

import meltygui_pycuda

The first release is 2026.1.post1, supporting Linux x86-64, CPython 3.12 and glibc 2.28 or newer. It installs meltygui_pycuda without overwriting the upstream pycuda package. Other platforms and Python versions are not supported yet.

OpenGL interoperability and cuRAND are enabled. Wheels bundle cuRAND with its NVIDIA CUDA 12.1 license; the driver and nvcc are not bundled. Runtime kernel compilation still requires a compatible CUDA toolkit and NVIDIA driver.

Build

Docker is required. The manylinux image and Python build dependencies are pinned. The wheel is built from the generated source archive, which is also published for local compilation when a matching wheel is unavailable.

python3 tools/build.py --cuda-root /usr/local/cuda-12.1
python3 tools/verify.py dist
uvx --from twine==7.0.0 twine check --strict dist/*.whl dist/*.tar.gz

Use --sudo-docker if Docker requires sudo. Source builds require a C++ compiler and a CUDA toolkit with nvcc on PATH.

Publish

Pushes to main and pull requests build and test without uploading. A v* tag or a manual run with publish=true publishes the verified wheel and source archive through PyPI trusted publishing. Published versions cannot be replaced.

Configure a pending GitHub publisher at https://pypi.org/manage/account/publishing/:

Then, after the main-branch build passes, create and push v2026.1.post1. No API token belongs in this repository.

Licenses and changes

See LICENSE, LICENSES, and MELTYGUI.md. The primary changes are the Python namespace, pinned build dependencies, Python 3.12 support, and OpenGL/cuRAND enabled by default.