Apache Arrow: The Great Library Unifier by Matt Topol, FOSDEM25

AI development on GPUs relies on a variety of low-level libraries — PyTorch, libcudf, TensorFlow, and others — each offering different strengths and specialized operations. Choosing between them can be difficult, especially when data movement between CPU and GPU becomes a performance bottleneck.

In this talk Matt Topol explains how the Apache Arrow data format and its C Device Interface, combined with DLPack, can enable seamless interoperability across these libraries while keeping data on the GPU and sharing device pointers.

We’ll walk through examples of transferring data between libraries without unnecessary GPU-CPU copies, using HuggingFace’s Arrow-based caches for training, and efficiently converting between Arrow and DLPack to unify multiple libraries into customized low-level AI pipelines.

All sessions at Low-level AI Engineering and Hacking, FOSDEM25 you can find here - https://archive.fosdem.org/2025/schedule/track/ai/

Видео Apache Arrow: The Great Library Unifier by Matt Topol, FOSDEM25 канала AIFoundry Org