Build Status

ONNXRuntime-Extensions is a library that extends the capability of the ONNX models and inference with ONNX Runtime, via the ONNX Runtime custom operator interface. It includes a set of Custom Operators to support common model pre and post-processing for audio, vision, text, and language models. As with ONNX Runtime, Extensions also supports multiple languages and platforms (Python on Windows/Linux/macOS, Android and iOS mobile platforms and Web assembly for web).

The basic workflow is to add the custom operators to an ONNX model and then to perform inference on the enhanced model with ONNX Runtime and ONNXRuntime-Extensions packages.

Pre and post-processing custom operators for vision, text, and NLP models This image was created using Combine.AI, which is powered by Bing Chat, Bing Image Creator, and EdgeGPT.


Python installation

pip install onnxruntime-extensions

Nightly Build

on Windows
pip install --index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/ORT-Nightly/pypi/simple/ onnxruntime-extensions

The onnxruntime-extensions package depends on onnx and onnxruntime.

on Linux/MacOS

Please make sure the compiler toolkit like gcc(later than g++ 8.0) or clang are installed before the following command

python -m pip install git+https://github.com/microsoft/onnxruntime-extensions.git

NuGet installation (with .NET CLI)

dotnet add package Microsoft.ML.OnnxRuntime.Extensions --version 0.8.1-alpha

iOS installation

In your CocoaPods Podfile, add the onnxruntime-extensions-c pod.


  # onnxruntime C/C++ full package
  pod 'onnxruntime-c'

  # onnxruntime-extensions C/C++ package
  pod 'onnxruntime-extensions-c'

Run pod install.

Android installation

In your Android Studio Project, make the following changes to:

  1. build.gradle (Project):

     repositories {
  2. build.gradle (Module):

     dependencies {
         // onnxruntime full package
         implementation 'com.microsoft.onnxruntime:onnxruntime-android:latest.release'
         // onnxruntime-extensions package
         implementation 'com.microsoft.onnxruntime:onnxruntime-extensions-android:latest.release'

Add pre and post-processing to the model

There are multiple ways to add pre and post processing to an ONNX graph:

If the pre processing operator is a HuggingFace tokenizer, you can also easily get the ONNX processing graph by converting from Huggingface transformer data processing classes such as in the following example:

import onnxruntime as _ort
from transformers import AutoTokenizer, GPT2Tokenizer
from onnxruntime_extensions import OrtPyFunction, gen_processing_models

# SentencePieceTokenizer
spm_hf_tokenizer = AutoTokenizer.from_pretrained("t5-base", model_max_length=512)
spm_onnx_model = OrtPyFunction(gen_processing_models(spm_hf_tokenizer, pre_kwargs={})[0])

# GPT2Tokenizer
gpt2_hf_tokenizer = GPT2Tokenizer.from_pretrained("Xenova/gpt-4", use_fast=False)
gpt2_onnx_model = OrtPyFunction(gen_processing_models(gpt2_hf_tokenizer, pre_kwargs={})[0])

For more information, you can check the API using the following:


What if I cannot find the custom operator I am looking for?

Find the custom operators we currently support here. If you do not find the custom operator you are looking for, you can add a new custom operator to ONNX Runtime Extensions like this. Note that if you do add a new operator, you will have to build from source.

Inference with ONNX Runtime and Extensions


There are individual packages for the following languages, please install it for the build.

import onnxruntime as _ort
from onnxruntime_extensions import get_library_path as _lib_path

so = _ort.SessionOptions()

# Run the ONNXRuntime Session as per ONNXRuntime docs suggestions.
sess = _ort.InferenceSession(model, so)
sess.run (...)


Register Extensions with a path to the Extensions shared library.

Ort::Env env = ...;

// Note: use `wchar_t` instead of `char` for paths on Windows
const char* model_uri = "/path/to/the/model.onnx";
const char* custom_op_library_filename = "/path/to/the/onnxruntime-extensions/shared/library";

Ort::SessionOptions session_options;

// Register Extensions custom ops with the session options.

// Create a session.
Ort::Session session(env, model_uri, session_options);

Register Extensions by calling the RegisterCustomOps function directly.

Ort::Env env = ...;

// Note: use `wchar_t` instead of `char` for paths on Windows
const char* model_uri = "/path/to/the/model.onnx";

Ort::SessionOptions session_options;

// Register Extensions custom ops with the session options.
// `RegisterCustomOps` is declared in onnxruntime_extensions.h.
Ort::ThrowOnError(RegisterCustomOps(static_cast<OrtSessionOptions*>(session_options), OrtGetApiBase()));

// Create a session.
Ort::Session session(env, model_uri, session_options);


var env = OrtEnvironment.getEnvironment();
var sess_opt = new OrtSession.SessionOptions();

/* Register the custom ops from onnxruntime-extensions */


SessionOptions options = new SessionOptions();
session = new InferenceSession(model, options);


Check out some end to end tutorials with our custom operators:


This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.microsoft.com.

When you submit a pull request, a CLA-bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., label, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.


MIT License

Table of contents