version: "1.0.0"

GammaGL Developer Skill

Use this skill when editing GammaGL source, examples, tests, installation metadata, or documentation.

Repository Layout

• gammagl/: library package.
• gammagl/layers/: reusable layers. Convolution layers live in

gammagl/layers/conv/.

• gammagl/models/: model definitions exported through

gammagl.models.__all__.

• gammagl/datasets/: dataset wrappers exported through

gammagl.datasets.__all__.

• gammagl/utils/: public utility functions exported through

gammagl.utils.__all__.

• examples/: runnable model examples and reproduction scripts.
• tests/: unit and smoke tests.
• setup.py, requirements.txt, docs/requirements.txt: package and

environment dependency entry points.

TensorLayerX Constraint

GammaGL relies on the GAMMA Lab TensorLayerX branch:

pip install git+https://github.com/dddg617/tensorlayerx.git@nightly

Install the backend first. For the default torch backend, install PyTorch before installing TensorLayerX. Do not replace this branch with the public PyPI TensorLayerX package when validating changes.

Quick Start Environments

CPU:

conda create -n gammagl-cpu python=3.10
conda activate gammagl-cpu
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
pip install git+https://github.com/dddg617/tensorlayerx.git@nightly
pip install pybind11 ninja
GAMMAGL_WITH_CUDA=0 pip install -e ".[build]" --no-build-isolation

CUDA:

conda create -n gammagl-cu python=3.10
conda activate gammagl-cu
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install git+https://github.com/dddg617/tensorlayerx.git@nightly
pip install pybind11 ninja
GAMMAGL_WITH_CUDA=auto pip install -e ".[build]" --no-build-isolation

Use one GammaGL package for CPU and GPU. Select CUDA extension compilation with GAMMAGL_WITH_CUDA=0, 1, or auto.

LLM/GFM features are optional:

pip install pybind11 ninja
pip install -e ".[llm-gfm]" --no-build-isolation

These extras may install packages such as PyTorch, Transformers, Torch-Geometric, OpenAI SDK, and Sentence Transformers. Core GammaGL imports must not require them.

Running Examples

Prefer TL_BACKEND=torch unless the task targets another backend.

TL_BACKEND=torch python examples/gcn/gcn_trainer.py --dataset cora --n_epoch 1
TL_BACKEND=torch python examples/gat/gat_trainer.py --dataset cora --n_epoch 1
TL_BACKEND=torch python examples/graphsage/train_full_trainer.py --dataset cora --n_epoch 1

If a trainer does not support a short epoch argument, first run --help and add a small smoke path instead of forcing a full experiment.

Public API Update Rules

When adding a layer:

• Add the implementation under the appropriate gammagl/layers/ subpackage.
• Export public classes from the subpackage __init__.py.
• If docs use from gammagl.layers import Name, export it in

gammagl/layers/__init__.py.

• Add or update a focused layer test.

When adding a model:

• Add the implementation under gammagl/models/.
• Export the public class from gammagl/models/__init__.py.
• Keep __all__ spelling and capitalization exactly aligned with imported

names.

• Add a focused model test or a tiny example smoke test.

When adding a dataset:

• Add the wrapper under gammagl/datasets/.
• Export it from gammagl/datasets/__init__.py.
• Avoid downloads during import. Download only when the dataset is constructed

or processed.

When adding a utility:

• Export public utilities from gammagl/utils/__init__.py.
• Do not import optional heavy frameworks at module import time.
• Add a unit test for shared behavior.

Dependency Rules

• Keep runtime dependencies small and backend-neutral.
• Keep pytest, ruff, pybind11, and ninja in development/build extras,

not the base runtime set.

• Keep molecular dependencies such as rdkit optional unless a core API truly

requires them.

• Keep LLM/GFM dependencies in extras. Do not import them from package

__init__ files.

• Do not add personal paths, hard-coded dataset roots, or import-time

TL_BACKEND overrides.

Troubleshooting

TensorLayerX import fails:

• Confirm PyTorch or the selected backend is installed first.
• Confirm TensorLayerX came from dddg617/tensorlayerx.git@nightly.

CUDA extension build fails:

• Use GAMMAGL_WITH_CUDA=0 for CPU-only installs.
• Use GAMMAGL_WITH_CUDA=auto when CUDA may or may not be present.
• Use GAMMAGL_WITH_CUDA=1 only when CUDA headers and nvcc are available.

Optional dependency is missing:

• Keep the error local to the optional feature.
• Add a clear installation hint for the relevant extra or package.
• Do not move the dependency into the core import path.

Dataset download fails:

• Keep tests on synthetic or cached tiny data where possible.
• Document dataset paths and credentials clearly in the example README.

Standard Checks

TL_BACKEND=torch python -m compileall -q gammagl tests examples
TL_BACKEND=torch python -c "from gammagl.layers import MessagePassing; from gammagl.models import *"
TL_BACKEND=torch python -m pytest tests/test_public_api.py -q