niro

Niro compiles computation graphs from ONNX (TensorFlow, PyTorch, JAX, and others planned) through one unified, strongly typed SSA IR. MLIR is the primary target; GIMPLE, QBE, and other backends may follow.

Read the documentation · IR specification · Python API

Warning

Niro is at an early stage and is not yet ready for production use. Pull requests may be rejected while the design is still taking shape.

Install

uv tool install git+https://github.com/cospectrum/niro.git

Usage

Compile a model to textual MLIR:

niro emit mlir model.onnx -o model.mlir

Input and output default to the standard streams, so Niro composes with other tools. Pass --input-format when there is no file extension to infer it from:

niro emit mlir --input-format onnx < model.onnx | mlir-opt

Inspect a model's entry-point signature:

niro inspect signature model.onnx

Contributing

See CONTRIBUTING.md.