Understanding API dispatch in Scientific Python eco-systems
summary
Claims that API dispatching speeds Python libraries without rewrites or heavy deps and shows it with concrete tools: backend discovery via Python entry_points (importlib.metadata), explicit conversion functions, and a small backend registry. A million-node centrality example dispatches NetworkX calls to a cuGraph GPU backend and yields orders-of-magnitude speedups for the subset of algorithms implemented. A parallel implementation for scikit-image routes image operations to compatible backends; NumPy uses a different, purely type-based dispatch.
AI-written summary · about
related talks
29mUnderstanding API Dispatching in NetworkX
27m100x Faster NetworkX: Dispatching to GraphBLAS
27mMultiple and Predicative Dispatch
45mPython for Speed, Scale, and Science
39mAccelerated Data Science Libraries
55mAccelerated Python: The Community and Ecosystem
37mThe Python Ecosystem for Data Science: A Guided Tour
28mNetworkX is Fast Now: Graph Analytics Unleashed