Zigrad

Zigrad is a programmable deep learning framework and compiler that aims to connect research workflows with production execution. Models are expressed in a structured program representation that can be inspected, transformed and lowered to different compiler backends without rewriting the model code. The design centers on a few ideas. The program representation is differentiable, with automatic differentiation sitting above the lowering path, and its semantics are backend agnostic, so the same program can be retargeted. Developers can write custom graph transformation passes to rewrite and optimize the program before backend lowering. The same program can run on backends such as XLA for large scale training or IREE for edge deployment. External kernel providers, including autotuning systems like TVM and the Mirage Superoptimizer, can generate kernels, and Zigrad can replace selected regions with specialized kernels only when profiling shows the change pays off. The site also offers an IR Viewer and an open explorer where users can read the terminology, run a demo, dump artifacts and inspect how programs change across lowering and execution boundaries. Documentation includes an API reference and a getting started guide, and the source is hosted on GitHub. Zigrad is for machine learning engineers, compiler developers and researchers who want control over how a model is compiled and optimized. The site does not list any pricing or paid plans.