dynG CUDA plugin (dyng-cu12, dyng-cu13)

This distribution adds the CUDA backends to dynG, the library that keeps the results of graph algorithms (shortest paths, multi-objective shortest paths, cycle counts) up to date while the graph changes in batches of edge insertions and deletions. It is a plugin of the dyng package and is installed through its extras:

pip install "dyng[cu13]"    # NVIDIA driver for CUDA 13 (580 or newer)
pip install "dyng[cu12]"    # NVIDIA driver for CUDA 12 (525 or newer)

dyng-cu12 is built with the CUDA 12 toolkit, dyng-cu13 with CUDA 13. Each contains one extension module with every backend of dynG (sequential, OpenMP and CUDA), compiled for the GPU architectures from Turing (sm_75) on, with the CUDA runtime linked in: no CUDA toolkit is needed at run time, only the NVIDIA driver. A plugin always has the same version as the dyng it belongs to, and pip installs that version of dyng with it.

import dyng

dyng.show_config()                       # which native module is active, and why
res = dyng.Resources.cuda(device=0)
g = dyng.Graph.from_edges([0, 0, 1, 2], [1, 2, 2, 3], [4, 1, 1, 5], resources=res)
tree = dyng.sssp.compute(g, source=0)    # runs on the GPU

You never import this package yourself: import dyng finds it through the entry point group dyng.backends and uses it when a CUDA driver and a device are present; otherwise dynG uses its CPU module and says why in dyng.show_config(). dyng.use_cpu_only() (before the first use) or DYNG_CPU_ONLY=1 (or true, yes, on) forces the CPU module.

Documentation: https://github.com/dyng-dev/dyng (installation, troubleshooting, the API). License: Apache-2.0 for dynG's own code; the wheel also contains third-party code (nanobind, robin-map, the GCC runtime, the CUDA runtime and the CUDA C++ Core Libraries) under the licences listed in its licence files (THIRD_PARTY_LICENSES.txt, THIRD_PARTY_LICENSES_CUDA.txt, NVIDIA_CUDA_EULA.txt).