ONNX Runtime v1.6.0

Announcements

• OpenMP will be disabled in future official builds (build option will still be available). A NoOpenMP version of ONNX Runtime is now available with this release on Nuget and PyPi for C/C++/C#/Python users.
• In the next release, MKL-ML, openblas, and jemallac build options will be removed, and the Microsoft.ML.OnnxRuntime.MKLML Nuget package will no longer be published. Users of MKL-ML are recommended to use the Intel EPs. If you are using these options and identify issues switching to an alternative build, please file an issue with details.

Key Feature Updates

General

• ONNX 1.8 support / opset 13
• New contrib ops: BiasSoftmax, MatMulIntegerToFloat, QLinearSigmoid, Trilu
• ORT Mobile now compatible with NNAPI for accelerating model execution on Android devices
• Build support for Mac with Apple Silicon (CPU only)
• New dependency: flatbuffers
• Support for loading sparse tensor initializers in pruned models
• Support for setting the execution priority of a node
• Support for selection of cuDNN conv algorithms
• BERT Model profiling tool

Performance

• New session option to disable denormal floating numbers on sse3 supporting CPUs
  • Eliminates unexpected performance degradation due to denormals without needing to retrain the model
• Option to share initializers between sessions to improve memory utilization
  • Useful when several models that use the same set of initializers except the last few layers of the model are loaded in the same process
  • Eliminates wasteful memory usage when every model (session) creates a separate instance of the same initializer
  • Exposed by the AddInitializer API
• Transformer model optimizations
  • Longformer: LongformerAttention CUDA operator added
  • Support for BERT models exported from Tensorflow with 1 or 2 inputs
  • Python optimizer supports additional models: openai-GPT, ALBERT and FlauBERT
• Quantization
  • Support of per-channel QuantizeLinear and DeQuantizeLinear
  • Support of LSTM quantization
  • Quantization performance improvement on ARM
  • CNN quantization perf optimizations, including u8s8 support and NHWC transformer in QLinearConv
• ThreadPool
  • Use _mm_pause() for spin loop to improve performance and power consumption

APIs and Packages

• Python - I/O Binding enhancements
  • Usage Documentation (OrtValue and IOBinding sections)
  • Python binding for the OrtValue data structure
    • An interface is exposed to allocate memory on a CUDA-supported device and define the contents of this memory. No longer need to use allocators provided by other libraries to allocate and manage CUDA memory to be used with ORT.
    • Allows consuming ORT allocated device memory as an OrtValue (check Scenario 4 in the IOBinding section of the documentation for an example)
  • OrtValue instances can be used to bind inputs/outputs. This is in addition to existing interfaces that allows binding a piece of memory directly/using numpy arrays that can be bound and may be particularly useful when binding ORT allocated device memory.
• C# - float16 and bfloat16 support
• Windows ML
  • NuGet package now supports UWP applications targeting Windows Store deployment for both CPU and GPU
  • Minor API Improvements:
    • Able to bind IIterable as inputs and outputs
    • Able to create Tensor* via multiple buffers
  • WindowsAI Redist now includes a statically linked C-Runtime package for additional deployment options

Execution Providers

• DNNL EP Updates
  • DNNL updated from 1.1.1 to 1.7
• NNAPI EP Updates
  • Support for CNN models
  • Additional operator support - Resize/Flatten/Clip
• TensorRT EP Updates
  • Int8 quantization support (experimental)
  • Engine cache refactoring and improvements
  • General fixes and performance improvements
• OpenVINO EP Updates
  • OpenVINO 2021.1 support
  • OpenVINO EP builds as shared library
  • Multi-threaded inferencing support
  • fp16 input type support
  • Multi-device plugin support
  • Hetero plugin support
  • Enable build on ARM64
• DirectML EP Updates (1.3.0 -> 1.4.0)
  • Utilizing the first public standalone release of the DirectML API through the DirectML NuGet package release
  • General fixes and improvements
• nGraph EP is removed. Recommend to use OpenVINO instead

Additional notes

• VCRuntime2019 with OpenMP: pinning a process to NUMA node 1 forces the execution to be single threaded. Fix is in progress in VC++.
  • Workaround: place the VS2017 vcomp DLL side-by-side so that ORT uses the VS2017 version
• Pip version >=20.3 is required for use on macOS Big Sur (11.x)
• The destructor of OrtEnv is now non-trivial and may do DLL unloading Do not call ReleaseEnv from DLLMain or put OrtEnv in global variables. It is not safe to call FreeLibrary from DllMain. - reference
• Some unit tests fail on Pascal GPUs. See: #5914
• If using the default CPU package (built with OpenMP), consider tuning the OpenMP settings to improve performance. By default the number of threads to use for openmp parallel regions is set to the number of logical CPUs. This may not be optimal for machines with hyper-threading; when CPUs are oversubscribed the 99-percentile latency could be 10x greater. Setting the OMP_NUM_THREADS environment variable to the number of physical cores is a good starting point. As noted in Announcements, future official builds of ORT will be published without OpenMP

Contributions

Contributors to ONNX Runtime include members across teams at Microsoft, along with our community members:
gwang-msft, snnn, skottmckay, edgchen1, hariharans29, wangyems, yufenglee, yuslepukhin, tianleiwu, SherlockNoMad, tracysh, ryanlai2, askhade, xadupre, liqunfu, RandySheriffH, jywu-msft, KeDengMS, pranavsharma, mrry, ashbhandare, iK1D, RyanUnderhill, MaajidKhan, wenbingl, kit1980, weixingzhang, tlh20, suffiank, Craigacp, smkarlap, stevenlix, zhanghuanrong, sfatimar, ytaous, tiagoshibata, fdwr, oliviajain, alberto-magni, jcwchen, mosdav, xzhu1900, wschin, codemzs, duli2012, smk2007, natke, zhijxu-MS, manashgoswami, zhangxiang1993, faxu, HectorSVC, take-cheeze, jingyanwangms, chilo-ms, YUNQIUGUO, jgbradley1, jessebenson, martinb35, Andrews548, souptc, pengwa, liuziyue, orilevari, BowenBao, thiagocrepaldi, jeffbloo

ONNX Runtime v1.5.3

This is a minor patch release on 1.5.2 with the following changes:

• Fix shared provider unload crash #5553
• Minor minimal build header fix

ONNX Runtime v1.5.2

This is a minor patch release on 1.5.1 with the following changes:

• Remove dependency on cudnn64_7.dll for GPU C# nuget: #5386
• Add config keys header file in the packages for Linux and Mac: #5388
• Add flatbuffers verifier for ORT format buffer: #5378
• Use official flatbuffers v1.12: #5392
• Mitigate pybind11 build break using Xcode 12 on macOS: #5381
• Support trilinear sampling in the Resize operator: #5300
• Update TensorRT parser to fix accuracy issue in some opset11 models: #5442

ONNX Runtime Training RC3.1

Fixes issue discovered during validation.

Changes:

• #5350

ONNX Runtime Training RC3

See: https://github.com/microsoft/onnxruntime/releases/tag/v1.5.1

ONNX Runtime v1.5.1

Key Updates

General

• Reduced Operator Kernel build allows ORT binaries to be built with only required operators in the model(s) - learn more
• [Preview] ORT for Mobile Platforms - minimizes build size for mobile and embedded devices - learn more
• Transformer model inferencing performance optimizations
  • Perf improvement for DistilBERT
  • Benchmark tool supports more pretrained models
• Improvements in quantization tool
  • Support quantization-aware training models
  • Make calibration tool to support general preprocessing and calibrate on input
  • Simplify the quantization APIs
  • Support of model larger than 2G
• New operators for static quantization: QLinearMul, QLinearAdd, QlinearSigmoid and QLinearLeakyRelu
• Prepack constant matrix B for float GEMM (MatMul, Attention)
• Limited Python 3.8 support added in addition to 3.5-3.7 for official Python packages. Not yet supported for Windows GPU and Linux ARM builds.
• Telemetry enabled in Java and NodeJS packages for Windows builds. Note: data is not directly sent to Microsoft or ORT teams by ONNX Runtime; enabling telemetry means trace events are collected by the Windows operating system and may be sent to the cloud based on the user's privacy settings - learn more.

API

• Python API support for RegisterCustomOpsLibrary
• IO Binding API for C/C++/C# language bindings. This allows use of pre-allocated buffers on targeted devices and also target device for unknown output shapes.
• Sharing of allocators between multiple sessions. This allows much better utilization of memory by not creating a separate arena for each session in the same process. See this for details.

Windows ML

• NuGet package now supports UWP applications targeting Windows Store deployment (CPU only)
• NuGet package now supports .NET and .NET framework applications
• RUST Developers can now deploy Windows ML – sample and documentation available here
• New APIs to for additional performance control:
  • IntraopNumThreads: Provides an ability to change the number of threads used in the threadpool for Intra Operator Execution for CPU operators through LearningModelSessionOptions.
  • SetNamedDimensionOverrides: Provides the ability to override named input dimensions to concrete values through LearningModelSessionOptions in order to achieve better runtime performance.
• Support for additional ONNX format image type denotations – Gray8, normalized [0..1] and normalized [-1..1]
• Reduced Windows ML package size by separating debug symbols into separate distribution package.

Execution Providers

• CUDA updates
  • CUDA 10.2 / cuDNN 8.0 in official package
  • CUDA 11 support added and available to build from source
  • CUDA conv kernel support asymmetrical padding to fully support models such as YoloV3 for improved GPU perf
• TensorRT EP updates
  • Support for TensorRT 7.1
  • Added TensorRT engine caching feature, turned on by setting env variable ORT_TENSORRT_ENGINE_CACHE_ENABLE=1
  • TensorRT builds are now built with the Execution Provider as a separate dll. If enabled in the build, the provider will be available as a shared library. This was previously also enabled for the DNNL EP (ORT 1.3). Other Execution Providers will be added in the future.
• OpenVINO EP updates
  • Support for OpenVINO 2020.4
  • Added runtime options for VPU hardware to select specific hardware device and enable fast compilation of models.
  • Enable C# binding support for OpenVINO EP
• DirectML EP updates
  • API available for Python (build from source) and C# Microsoft.ML.OnnxRuntime.DirectML
  • 7 new operators for ONNX 1.7 (opset 12): Celu, GreaterOrEqual, LessOrEqual, ArgMin/Max with select_last_index, GatherND with batch_dim, RoiAlign
  • New data integer types were added to existing operators: Clip int, Max int, Min int, MaxPool int8, ReduceMin int8, ReduceMax int8, Pow int exponent
  • Higher dimension support 1D to 8D added to these operators: ElementWise*, Activation*, Reduce*, ArgMin/ArgMax, Gather*, Scatter*, OneHot
  • 64-bit support for indices on GPU's that support it: Gather, Scatter, OneHot, ArgMax/ArgMin, Cast.
• Android NNAPI EP updates:
  • Support for dynamic input shape
  • Int32/float32/uint8 data type
  • 50% more supported operators (36 total)
  • Support for Uint8 static quantization
  • Smaller binary size
  • Lower memory consumption
  • CPU fallback for Android level 26-
• MiGraphX EP updates
  • Added ONNX operators: GatherElements, NonZero, Equal, and Where
  • Support for Boolean data type
  • Improve support for existing operators:
    • Asymmetric padding of AveragePool
    • Multi-dimensional support for Convolution, Pooling, LRN, and Batchnormalization
    • Ceil mode support for AveragePool and MaxPool
    • More general approach to check whether constant folding is possible
  • Improved graph partitioning logic

Training (RC3 release)

• New and improved API to simplify integration with PyTorch trainer code - see instructions here
• Updated CUDA 11 / cuDNN 8.0 support to accelerate in NVIDIA A100

Dependency updates

MacOS binaries now rely on openmp to be installed. See this for reference.

Contributions

Contributors to ONNX Runtime include members across teams at Microsoft, along with our community members:

gwang-msft, snnn, skottmckay, hariharans29, thiagocrepaldi, tianleiwu, wangyems, RandySheriffH, yufenglee, SherlockNoMad, smk2007, jywu-msft, liqunfu, edgchen1, yuslepukhin, tiagoshibata, fdwr, ashbhandare, iK1D, wschin, BowenBao, zhanghuanrong, RyanUnderhill, ryanlai2, askhade, pranavsharma, martinb35, suffiank, ytaous, KeDengMS, rayankrish, natke, YUNQIUGUO, range4life, smkarlap, zhangxiang1993, xzhu1900, codemzs, weixingzhang, stevenlix, tracysh, mosdav, jingyanwangms, tlh20, souptc, orilevari, kit1980, yangchen-MS, faxu, fs-eire, wenbingl, chilo-ms, xkszltl, Andrews548, yuzawa-san, MaximKalininMS, jgbradley1, nickfeeney, zhijxu-MS, Tixxx, suryasidd, Craigacp, duli2012, jeffbloo

ORTTraining RC2

orttraining_rc2

For ORTTraining RC2 release to include hotfix https://github.com/micr…

ONNX Runtime v1.4.0

Key Updates

• Performance optimizations for Transformer models
  • GPT2 - Enable optimizations for Attention with Past State and Attention Mask
  • BERT - Improve EmbedLayerNormalization fusion coverage
• Quantization updates
  • Added new quantization operators: QLinearAdd, QAttention
  • Improved quantization performance for transformer based models on CPU
    • More graph fusion
    • Further optimization in MLAS kernel
    • Introduced pre-packing for constant Matrix B of DynamicQuantizeMatMul and Qattention
• New Python IOBinding APIs (bind_cpu_input, bind_output, copy_outputs_to_cpu) allow easier benchmarking
  • Users no longer need to allocate inputs and outputs on non-CPU devices using third-party allocators.
  • Users no longer need to copy inputs to non-CPU devices; ORT handles the copy.
  • Users can now use copy_outputs_to_cpu to copy outputs from non-CPU devices to CPU for verification.
• CUDA support for Einsum (opset12)
• ONNX Runtime Training updates
  • Opset 12 support
  • New sample for training experiment using Huggingface GPT-2.
    • Upgraded docker image built from the latest PyTorch release
• Telemetry is now enabled by default for Python packages and Github release zip files (C API); see more details on what/how telemetry is collected in ORT
• [Coming soon] Availability of Python package for ONNX Runtime 1.4 for Jetpack 4.4

Execution Providers

New Execution Providers available for preview:

• [Preview] AMD MIGraphX
• [Preview] ARM NN

Contributions

Contributors to ONNX Runtime include members across teams at Microsoft, along with our community members:

snnn, tianleiwu, edgchen1, hariharans29, skottmckay, tracysh, yufenglee, fs-eire, codemzs, tiagoshibata, yuslepukhin, gwang-msft, wschin, smk2007, prabhat00155, liuziyue, liqunfu, ytaous, iK1D, BowenBao, askhade, pranavsharma, faxu, jywu-msft, ryanlai2, xzhu1900, KeDengMS, tlh20, smkarlap, weixingzhang, jeffbloo, RyanUnderhill, mrry, jgbradley1, stevenlix, zhanghuanrong, suffiank, Andrews548, pengwa, SherlockNoMad, orilevari, duli2012, yangchen-MS, yan12125, jornt-xilinx, ashbhandare, neginraoof, Tixxx, thiagocrepaldi, Craigacp, mayeut, chilo-ms, prasanthpul, martinb35, manashgoswami, zhangxiang1993, suryasidd, wangyems, kit1980, RandySheriffH, fdwr

ONNX Runtime v1.3.1

This update includes changes to support the published packages for the Java and nodejs APIs for the 1.3.0 release.

• Maven: Java API CPU
• Maven: Java API GPU
• NPM: ONNX Runtime Node.js API

For all other APIs/builds, the 1.3.0 release packages are suggested. 1.3.1 does address the 1.3.0 issue of Crash when setting IntraOpNumThreads using the C/C++/C# API, so if this fix is needed it can be built from source using this release branch (with official release support).

ONNX Runtime v1.3.0

Key Updates

General

• ONNX 1.7 support
  • Opset 12
  • Function expansion support that enables several new ONNX 1.7 ops such as NegativeLogLikelihoodLoss, GreaterOrEqual, LessOrEqual, Celu to run without a kernel implementation.
• [Preview] ONNX Runtime Training
  • ONNX Runtime Training is a new capability released in preview to accelerate training transformer models. See the sample here to use this feature in your training experiments.
• Improved threadpool support for better resource utilization
  • Improved threadpool abstractions that switch between openmp and Eigen threadpools based on build settings. All operators have been updated to use these new abstractions.
  • Improved Eigen based threadpool now allow ops to provide cost (among other things like thread affinity) for operations
  • Simpler configuration of thread count. If built with OpenMP, use the OpenMP env variables; else use the ORT APIs to configure the number of threads.
  • Support for sessions to share global threadpool. See this for more information.
• Performance improvements
  • ~10% average measured latency improvements amongst key representative models (including ONNX model zoo models, MLPerf, and production models shipped in Microsoft products)
  • Further latency improvements for Transformer models on CPU and GPU - benchmark script
  • Improved batch inferencing latency for scikit-learn models for large batch sizes
    • Significant improvements in the implementations of the following ONNX operators: TreeEnsembleRegressor, TreeEnsembleClassifier, LinearRegressor, LinearClassifier, SVMRegressor, SVMClassifier, TopK
  • C# API optimizations - PR3171
• Telemetry enabled for Windows (more details on telemetry collection)
• Improved error reporting when a kernel cannot be found due to missing type implementation
• Minor fixes based on static code analysis

Dependency updates

Please note that this version of onnxruntime depends on Visual C++ 2019 runtime. Previous versions depended on Visual C++ 2017. Please also refer https://github.com/microsoft/onnxruntime/tree/rel-1.3.0#system-requirements for the full set of system requirements.

APIs and Packages

• [General Availability] Windows Machine Learning APIs - package published on Nuget - Microsoft.AI.MachineLearning
  • Performance improvements
  • Opset updates
• [General Availability] ONNX Runtime with DirectML package published on Nuget -Microsoft.ML.OnnxRuntime.DirectML
• [General Availability] Java API - Maven package coming soon.
• [Preview] Javascript (node.js) API now available to build from the master branch.
• ARM64 Linux CPU Python package now available on Pypi. Note: this requires building ONNX for ARM64.
• Nightly dev builds from master (Nuget feed, TestPypi-CPU, GPU)
• API Updates
  • I/O binding support for Python API - This reduces execution time significantly by allowing users to setup inputs/outputs on the GPU prior to model execution.
  • API to specify free dimensions based on both denotations and symbolic names.

Execution Providers

• OpenVINO v2.0 EP
• DirectML EP updates
  • Updated graph interface to abstract GPU-dependent graph optimization
  • ONNX opset 10 and 11 support
  • Initial support of 8bit and quantized operators
  • Performance optimizations
• [Preview] Rockchip NPU EP
• [Preview] Xilinx FPGA Vitis-AI EP
• Capability to build execution providers as DLLs - supported for DNNL EP, work in progress for other EPs.
  • If enabled in the build, the provider will be available as a shared library. Previously, EPs had to be statically linked with the core code.
  • No runtime cost to include the EP if it isn't loaded; can now dynamically decide when to load it based on the model

Contributions

We'd like to recognize our community members across various teams at Microsoft and other companies for all their valuable contributions. Our community contributors in this release include: Adam Pocock, pranavm-nvidia, Andrew Kane, Takeshi Watanabe, Jianhao Zhang, Colin Jermain, Andrews548, Jan Scholz, Pranav Prakash, suryasidd, and S. Manohar Karlapalem.

The ONNX Runtime Training code was originally developed internally at Microsoft, before being ported to Github. We’d like to recognize the original contributors: Aishwarya Bhandare, Ashwin Kumar, Cheng Tang, Du Li, Edward Chen, Ethan Tao, Fanny Nina Paravecino, Ganesan Ramalingam, Harshitha Parnandi Venkata, Jesse Benson, Jorgen Thelin, Ke Deng, Liqun Fu, Li-Wen Chang, Peng Wang, Sergii Dymchenko, Sherlock Huang, Stuart Schaefer, Tao Qin, Thiago Crepaldi, Tianju Xu, Weichun Wang, Wei Zuo, Wei-Sheng Chin, Weixing Zhang, Xiaowan Dong, Xueyun Zhu, Zeeshan Siddiqui, and Zixuan Jiang.

Known Issues

1. The source doesn't compile on Ubuntu 14.04. See #4048
2. Crash when setting IntraOpNumThreads using the C/C++/C# API. Fix is available in the master branch.
   Workaround: Setting IntraOpNumThreads is inconsequential when using ORT that is built with openmp enabled. Hence it's not required and can be safely commented out. Use the openmp env variables to set the threading params for openmp enabled builds (which is the recommended way).