Build WholeGraph from source#

WholeGraph is built from the cuGraph-GNN repository. The default build installs both libwholegraph and pylibwholegraph.

Prerequisites#

Use a Linux system and CUDA version supported by RAPIDS 26.10. The generated development environment in the repository is the authoritative dependency list.

git clone https://github.com/rapidsai/cugraph-gnn.git
cd cugraph-gnn

conda env create \
  --name cugraph_gnn_dev \
  --file conda/environments/all_cuda-133_arch-x86_64.yaml
conda activate cugraph_gnn_dev

Choose the generated environment file that matches the desired supported CUDA and architecture combination when more than one is present.

Build the native and Python libraries#

./build.sh clean
./build.sh libwholegraph pylibwholegraph

The build script installs into $PREFIX when it is set, then $CONDA_PREFIX, and otherwise the system prefix. Useful options include:

  • -n to build without installing;

  • -v for verbose output;

  • -g for a debug build;

  • --pydevelop for an editable Python install;

  • --allgpuarch to build all supported GPU architectures.

Run ./build.sh --help for the current complete option list.

Run tests#

Build the C++ test targets with:

./build.sh tests

Run the Python tests from the repository root:

python -m pytest python/pylibwholegraph

See the cuGraph-GNN contributing guide for developer workflow and style requirements.