NVIDIA GPUs are becoming a key tool for accelerated computing, with many Python libraries in ML and HPC eager to harness their power. Open-source projects face a major challenge: with contributions from around the world, how can we ensure that the specialized code for GPUs remains functional and reliable?
The XGBoost project (Python ML library) tackled this challenge head-on. In this poster, we will share our experience setting up a CI pipeline for building and testing native code targeting NVIDIA GPUs. Additionally, we will offer practical recommendations for other open-source developers looking to adopt GPUs in their projects. Special focus will be given to common constraints in open-source development, such as budget, developer time, development velocity, and domain knowledge. We will also provide helpful pointers for packaging the native code as Python wheels and publishing them to PyPI and Conda-forge.