Scikit-learn on GPUs with Array API

on YouTube ↗36m

" summary

Scikit-learn 1.3 runs a subset of estimators on accelerators by dispatching array operations to CuPy and PyTorch via the Array API standard, yielding substantial speedups on GPUs and even faster CPU kernels when using PyTorch backends. Support is incremental (only some estimators), compiled SciPy C code currently falls back by converting to NumPy, and a compatibility shim array_api_compat plus experimental plugin dispatch enable wider device support but require extra work to handle API and semantic differences.

AI-written summary · about

" chapters (6)

1. 0:00GPU support in scikit-learn
2. 4:00Array API user-facing API
3. 9:29Array API standard
4. 18:24Migrating to Array API
5. 31:43Compiled code with Array API
6. 35:40Adopt Array API now

" related talks

• 28mA Hitchhiker's Guide to the Array API Standard EcosystemLucas Colley · PyData Paris 2025
• 05mKeynote: The Promise of PyTorch as a General-Purpose Array-Oriented Computational..Travis E. Oliphant · PyTorch Conference 2023
• 30mCuPy: A NumPy-compatible Library for GPUShohei Hido · PyCon 2018
• 25mScale Smarter, Not Harder, with cuPyNumeric.Irina Demeshko, Quynh L. Nguyen · PyCon US 2025
• 02h 28mGPU Accelerated PythonMcCarty, Riehl, Tomlinson · PyData NYC 2024
• 24mZero Code Change Acceleration: familiar interfaces and high performanceTim Head · PyCon DE & PyData 2025
• 47mBig ideas shaping scientific Python: the quest for performance and usabilityRalf Gommers · PyData Paris 2025
• 02h 55mGPU Accelerated Python w: CUDAJacob Tomlinson · PyData London 2025