Repolode
Submit
onnx

onnx/onnx-mlir

A compiler that lowers ONNX neural network graphs through MLIR into code with minimal runtime dependencies.

View on GitHub
Stars
1,066
Forks
457
Open issues
357
Last commit
yesterday

What is it?

ONNX-MLIR takes a valid ONNX model graph and compiles it into code that implements the graph, aiming for minimal runtime support. It implements the ONNX standard and is built on the LLVM/MLIR compiler infrastructure.

The project provides an ONNX dialect that other MLIR-based projects can integrate, compiler interfaces that lower ONNX graphs to MLIR files, LLVM bytecode, or C and Java libraries, and an onnx-mlir command-line driver. It also includes Python, C, C++, and Java runtime interfaces for loading and running compiled models.

Support for individual ONNX operations is tracked per target, including a generic CPU backend and IBM's Telum integrated AI accelerator. Builds run natively on Linux, macOS, and Windows, and Docker images are offered as the recommended setup.

Key features

  • 01Converts ONNX graphs into MLIR, LLVM IR, object files, shared libraries, or JAR files.
  • 02Provides an ONNX dialect that other MLIR projects can reuse.
  • 03Offers Python, C, C++, and Java runtime interfaces.
  • 04Supports optimization levels O0 through O3 and can read flags from an environment variable.
  • 05Publishes per-operation support tables for CPU and IBM Telum NNPA targets.

Use cases

  • Deploying ONNX models as compiled libraries with minimal runtime dependencies.
  • Targeting IBM Telum servers with ONNX models via the NNPA backend.
  • Building compiler tooling that integrates the ONNX dialect into an MLIR pipeline.

Quick start

  1. 1

    Set up with prebuilt Docker containers (see docs/Docker.md)

  2. 2

    Build ONNX-MLIR directly on Linux or macOS (see docs/BuildOnLinuxOSX.md)

  3. 3

    Lower an ONNX model to the ONNX dialect with the built driver

    ./onnx-mlir --EmitONNXIR add.onnx

Things to know

  • Building from source requires a specific LLVM/MLIR commit and several prerequisites such as clang, protobuf and cmake.
  • There is no simple package-manager install; setup can be tricky on some systems, so Docker is recommended.
  • Operation support is limited to the operations listed in the project's supported-ops documents.

In collections

Compilers, transpilers & programming languages

Topics

#onnx#mlir#compiler#deep-learning#llvm#inference#machine-learning

FAQ

What is onnx-mlir?
A compiler that lowers ONNX neural network graphs through MLIR into code with minimal runtime dependencies. ONNX-MLIR takes a valid ONNX model graph and compiles it into code that implements the graph, aiming for minimal runtime support. It implements the ONNX standard and is built on the LLVM/MLIR compiler infrastructure.
Is onnx-mlir free to use? What license does it have?
onnx-mlir is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is onnx-mlir still maintained?
Actively maintained. The last commit to onnx-mlir was on Nov 6, 2026.
What can I use onnx-mlir for?
Deploying ONNX models as compiled libraries with minimal runtime dependencies. Targeting IBM Telum servers with ONNX models via the NNPA backend. Building compiler tooling that integrates the ONNX dialect into an MLIR pipeline.
How do I install onnx-mlir?
Set up with prebuilt Docker containers (see docs/Docker.md) Build ONNX-MLIR directly on Linux or macOS (see docs/BuildOnLinuxOSX.md) Lower an ONNX model to the ONNX dialect with the built driver: ./onnx-mlir --EmitONNXIR add.onnx
Where can I get onnx-mlir?
The source code and installation instructions are on GitHub: https://github.com/onnx/onnx-mlir

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects