ONNX Runtime is an active, fast-paced project backed by a strong team of Microsoft engineers and data scientists along with a worldwide community of partners and contributors. This roadmap summarizes the pending investments identified by the team to continually grow ONNX Runtime as a robust, versatile, and high performance inference engine for DNN and traditional ML models.
ONNX Runtime fully aligns with the goals of the ONNX community to bring openness and interoperability to the AI industry. ONNX is a standardization specification, and ONNX Runtime is an inference engine for models in the ONNX format to efficiently productionize models with high performance. For key technical design objectives and considerations, please see ONNX Runtime High Level Design.
We recognize the challenges involved in operationalizing ML models performantly in an agile way, and we understand that high volume production services can be highly performance-sensitive and often need to support a variety of compute targets (we experience these first-hand at Microsoft across our vast array of products and services).
As such, our investments are directly in support of solving those challenges, focusing on areas such as:
- Platform coverage
- Extensibility and customization
- Performance (latency, memory, throughput, scale, etc)
- Model coverage
- Quality and ease of use - including backwards compatibility of models (older opsets) and APIs
In addition to our OSS participation, we also internally use this technology in core products at Microsoft, with over 60 models in production providing an average of 2x+ performance improvement.
In support of the high level goals outlined above, the investment areas listed below represent our active and backlog projects, which are largely driven by community demand and anticipated usage opportunities. We will work through our prioritized backlog as quickly as possible, and if there are any specific features or enhancements you need, we gladly welcome community contributions for these efforts or any of the enhancements suggested on Github. If you have a specific suggestion or unsupported use case, please let us know by filing a Github issue.
Platform coverage
Extensibility and customization
Performance
Model coverage
Quality and ease of use
ONNX Runtime already supports a wide range of architectures, platforms, and languages, and this will continue to be an active investment area to broaden the availability of the engine for varied usage. Additionally, we understand that lightweight devices and local applications may have constraints for package size, so there is active awareness to opportunistically minimize binary size.
| Supported | Future |
|---|---|
| X64 | Additional as-needed |
| X86 | |
| ARM64 (Limited) | |
| ARM32 (Limited) |
| Supported | Future |
|---|---|
| Windows 7+ | iOS |
| Linux (various) | |
| Mac OS X | |
| Android (community contribution, Preview) |
| Supported | Future |
|---|---|
| Python (3.5, 3.6, 3.7) | Javascript |
| C++ | |
| C# | |
| C | |
| Ruby (community project) | |
| Java |
To achieve the best performance on a growing set of compute targets across cloud and the intelligent edge, we invest in and partner with hardware partners and community members to add new execution providers. The flexible pluggability of ONNX Runtime is critical to support a broad range of scenarios and compute options.
| Supported | Future |
|---|---|
| MLAS (Microsoft Linear Algebra Subprograms) | AMD GPU |
| Intel DNNL / MKL-ML | Xilinx FPGA |
| Intel nGraph | |
| NVIDIA CUDA | |
| NVIDIA TensorRT | |
| Intel OpenVINO | |
| Nuphar Model Compiler | |
| Microsoft Direct ML | |
| Android NN API (Preview) | |
| ARM Compute Library (Preview) |
To maximize performance potential, we will be continually adding additional CUDA implementations for supported operators.
In addition to new execution providers, we aim to make it easy for community partners to contribute in a non-disruptive way. To support this, we are investing in improvements to the execution provider interface for easily registering new execution providers and separating out EPs from the core runtime engine.
Performance is a key focus for ONNX Runtime. From latency to memory utilization to CPU usage, we are constantly seeking strategies to deliver the best performance. Although DNNs are rapidly driving research areas for innovation, we acknowledge that in practice, many companies and developers are still using traditional ML frameworks for reasons ranging from expertise to privacy to legality. As such, ONNX Runtime is focused on improvements and support for both DNNs and traditional ML.
- Improvements to batch processing for scikit-learn models
- More quantization support
- Improved multithreading (e.g. smarter work sharding, consolidation of different types of thread pools, user supplied thread pools, etc)
- Graph optimizations
- Intelligent graph partitioning to maximize the value of different accelerators
IoT provides growing opportunity to execute ML workloads on the edge of the network, where the data is collected. However, the devices used for ML execution have different hardware specifications. To support compatibility with this group of devices, we will invest in strategies to optimize ONNX model execution across the breadth of IoT endpoints using different hardware configurations with CPUs, GPUs and custom NN ASICs.
The ONNX spec focuses on ML model interoperability rather than coverage of all operators from all frameworks. We aim to continuously improve coverage to support popular as well as new state-of-the-art models.
As more operators are added to the ONNX spec, ONNX Runtime will provide implementations (default CPU and GPU-CUDA) of each to stay in compliance with the latest ONNX spec.
A few specific items include:
- Sparse Tensor support
- Generic function logic without separate kernels
We work with the OSS and ONNX community to ensure popular frameworks can export or be converted to ONNX format.
- PyTorch export
- Tensorflow-ONNX
- Keras-ONNX
- Sklearn-ONNX
- ONNXMLTools (CoreML, XGBoost, LibSVM, LightGBM, SparkML)
- ML.NET
To decrease the risk of model inferencing failures, we will improve the error handling and fallback strategies for missing types or unsupported operators. For EPs that have missing or incorrect implementations for ONNX operators, we aim to fallback or fail as gracefully as possible.
Focusing on practicality, we take a scenario driven approach to adding additional capabilities to ONNX Runtime.
We understand that data scientists and ML engineers work with many different products and toolsets to bring complex machine learning algorithms to life through innovative user-facing applications. We want to ensure ONNX Runtime works as seamlessly as possible with these. If you've identified any integration ideas or opportunities and have questions or need assistance, we encourage use of Github Issues as a discussion forum.
Some of these products include:
- AzureML: simplify the process to train, convert, and deploy ONNX models to Azure
- Model Interpretability: explainability for ONNX models
- ML.NET: inference ONNX models in .NET
- PyTorch: improve coverage for exporting trained models to ONNX
- Windows: run ONNX models on Windows devices using the built-in Windows ML APIs. Windows ML APIs will be included in the ONNX Runtime builds and binaries to enable Windows developers to get OS-independent updates
- SQL Database Edge: predict with ONNX models in SQL Database Edge, an optimized relational database engine geared for IoT and IoT Edge deployments
Have an idea or feature request? Contribute or let us know!