Zigrad
Zigrad is a programmable deep learning framework and compiler that bridges research workflows and production execution.
Programmable program representation
One program representation, multiple compilation paths.
Models are represented in a structured program representation that can be inspected, transformed, and lowered to different compiler backends without rewriting model code.
Programmable graph transformations
Apply custom passes to rewrite and optimize the program representation before backend lowering.
Differentiable program representation
Automatic differentiation is abstracted above the lowering and compilation path.
Backend-agnostic semantics
The program representation encodes domain-level operations without assuming a compilation target, enabling retargeting without rewriting model code.
Retargeting and selective kernelization
Retarget the same program and specialize only the regions that matter.
The same program can target different compiler backends while selectively replacing regions with specialized kernels when profitability warrants it.
Backend retargeting
Execute the same program across compiler backends such as XLA for large-scale training or IREE edge deployment.
External kernel providers
Generate kernels with autotuning systems such as TVM or the Mirage Superoptimizer.
Profitability-gated specialization
Identify candidate regions via pattern matching or compiler hints and rewrite them with specialized kernels only when profiling justifies it.
Inspect the pipeline end to end.
Read the terminology, run a demo, dump the artifacts, and inspect how programs change across lowering and execution boundaries.