Zero Code Change Acceleration: familiar interfaces and high performance
summary
Demonstrates keeping standard pandas/scikit-learn code while switching implementations under the hood: an import-hook that replaces modules with wrappers that choose between scikit-learn and NVIDIA QML at fit time, plus rewrites using the array API compatibility so PyTorch tensors sit on the GPU. Requires one magic cell or two top-line imports, accelerates some workloads 10–20×, covers a growing subset of algorithms (metrics easy, estimators harder) and is available experimentally since 2023.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:10Python REPL drives fast iteration
- 3:00Consistent data science APIs
- 4:26Cost of slow analysis
- 5:48Zero code change strategies
- 6:51Modifying import with QML.XL
- 8:18Python import system
- 9:24Finder and loader design
- 11:50Performance gains with QML.XL
- 12:54scikit-learn GPU rewrite
- 13:58scikit-learn's GPU stance
- 15:24Multiple array APIs challenge
- 16:29Array API standard
- 17:53scikit-learn Array API progress
- 19:00Familiar APIs vs performance
- 20:30Q&A on array conversion
- 23:25Final Q&A
related talks
39mAccelerated Data Science Libraries
55mAccelerated Python: The Community and Ecosystem
23mHugging Face Accelerate: Making Device-Agnostic ML Training and Inference Easy...
29mcudf.pandas: The Zero Code Change GPU Accelerator for Pandas
35mPython at the Speed of Light: A Complete Tour of the CUDA Ecosystem
24mAdvancements in High-Performance AI/ML through PyTorch's Python Compiler
27mAccelerating Generative AI
30mSnap ML: Accelerated, Accurate,Efficient,Machine Learning