weld

Why we included this project

Most analytics pipelines spend more time shuffling data between libraries than actually computing. Weld tackles that by building a lazy computation graph for the whole workflow and compiling it to native code with LLVM, so the optimizer sees across function boundaries instead of treating each call as a black box. If you're stuck with pandas or NumPy code that's slow because of cross-library data movement rather than single-function compute, this is worth a look. The repo includes Grizzly, a Pandas-on-Weld subset, and a Weld-ified NumPy, which help you try the runtime against familiar APIs. Just know it's a research-grade project from Stanford, so plan on building and integrating it yourself rather than expecting a drop-in accelerator.