26 KiB
| title | description | parent | nav_order | redirect_from |
|---|---|---|---|---|
| Intel - OpenVINO™ | Instructions to execute OpenVINO™ Execution Provider for ONNX Runtime. | Execution Providers | 3 | /docs/reference/execution-providers/OpenVINO-ExecutionProvider |
OpenVINO™ Execution Provider
{: .no_toc }
Accelerate ONNX models on Intel CPUs, GPUs and VPUs with Intel OpenVINO™ Execution Provider. Please refer to this page for details on the Intel hardware supported.
Contents
{: .no_toc }
- TOC placeholder {:toc}
Install
Pre-built packages and Docker images are published for OpenVINO™ Execution Provider for ONNX Runtime by Intel for each release.
- OpenVINO™ Execution Provider for ONNX Runtime Release page: Latest v4.3 Release
- OpenVINO™ Execution Provider for ONNX Runtime Release page: Latest v4.2 Release
- Python wheels Ubuntu/Windows: onnxruntime-openvino
- Docker image: openvino/onnxruntime_ep_ubuntu20
Requirements
ONNX Runtime OpenVINO™ Execution Provider is compatible with three lastest releases of OpenVINO™.
| ONNX Runtime | OpenVINO™ | Notes |
|---|---|---|
| 1.13.0 | 2022.3 | Details |
| 1.13.0 | 2022.2 | Details |
| 1.11.0 | 2022.1 | Details |
| 1.10.0 | 2021.4.2 | Details |
| 1.9.0 | 2021.4.1 | Details |
| 1.8.1 | 2021.4 | Details |
| 1.8.0 | 2021.3 | Details |
Build
For build instructions, please see the BUILD page.
Usage
OpenVINO™ Environment
Python
Please download onnxruntime-openvino python packages from PyPi.org:
pip install onnxruntime-openvino==1.13.0
To enable OpenVINO™ Execution Provider with ONNX Runtime on Windows we must install OpenVINO™ separately:
pip install openvino==2022.3
Code to be added in ONNX Runtime Windows Samples:
import openvino.utils as utils
utils.add_openvino_libs_to_path()
OpenVINO™ Execution Provider with Onnx Runtime on Linux installed from PyPi.org come with prebuilt OpenVINO™ libs and supports flag CXX11_ABI=0. So there is no need to install OpenVINO™ separately.
To enable CX11_ABI=1 flag, build Onnx Runtime python wheel packages from source. For build instructions, please see the BUILD page. OpenVINO™ Execution Provider wheels on Linux built from source will not have prebuilt OpenVINO™ libs so we must set the OpenVINO™ Environment Variable using the full installer package of OpenVINO™:
C:\ <openvino_install_directory>\setupvars.bat
C++
For Running C++/C# ORT Samples with the OpenVINO™ Execution Provider it is must to set up the OpenVINO™ Environment Variables using the full installer package of OpenVINO™. Initialize the OpenVINO™ environment by running the setupvars script as shown below. This is a required step:
- For Linux run:
$ source <openvino_install_directory>/setupvars.sh
- For Windows run:
C:\ <openvino_install_directory>\setupvars.bat
Note: If you are using a dockerfile to use OpenVINO™ Execution Provider, sourcing OpenVINO™ won't be possible within the dockerfile. You would have to explicitly set the LD_LIBRARY_PATH to point to OpenVINO™ libraries location. Refer our dockerfile.
C#
To use csharp api for openvino execution provider create a custom nuget package. Follow the instructions here to install prerequisites for nuget creation. Once prerequisites are installed follow the instructions to build openvino execution provider and add an extra flag --build_nuget to create nuget packages. Two nuget packages will be created Microsoft.ML.OnnxRuntime.Managed and Microsoft.ML.OnnxRuntime.Openvino.
Multi-threading for OpenVINO™ Execution Provider
OpenVINO™ Execution Provider for ONNX Runtime enables thread-safe deep learning inference
Heterogeneous Execution for OpenVINO™ Execution Provider
The heterogeneous execution enables computing for inference on one network on several devices. Purposes to execute networks in heterogeneous mode:
- To utilize accelerator's power and calculate the heaviest parts of the network on the accelerator and execute unsupported layers on fallback devices like the CPU to utilize all available hardware more efficiently during one inference.
For more information on Heterogeneous plugin of OpenVINO™, please refer to the Intel OpenVINO™ Heterogeneous Plugin.
Multi-Device Execution for OpenVINO EP
Multi-Device plugin automatically assigns inference requests to available computational devices to execute the requests in parallel. Potential gains are as follows:
- Improved throughput that multiple devices can deliver (compared to single-device execution)
- More consistent performance, since the devices can now share the inference burden (so that if one device is becoming too busy, another device can take more of the load)
For more information on Multi-Device plugin of OpenVINO™, please refer to the Intel OpenVINO™ Multi Device Plugin.
Auto-Device Execution for OpenVINO EP
Use AUTO:<device 1><device 2>.. as the device name to delegate selection of an actual accelerator to OpenVINO™. Auto-device internally recognizes and selects devices from CPU, integrated GPU and discrete Intel GPUs (when available) depending on the device capabilities and the characteristic of CNN models, for example, precisions. Then Auto-device assigns inference requests to the selected device.
From the application point of view, this is just another device that handles all accelerators in full system.
For more information on Auto-Device plugin of OpenVINO™, please refer to the Intel OpenVINO™ Auto Device Plugin.
Model caching feature for OpenVINO™ Execution Provider
The model caching setting enables blobs with CPU and iGPU.
OpenCL queue throttling for GPU devices
Enables OpenCL queue throttling for GPU devices. Reduces CPU utilization when using GPUs with OpenVINO EP.
Model caching
OpenVINO™ supports model caching.
From OpenVINO™ 2022.1 version, model caching feature is supported on Myriadx(VPU), CPU and kernel caching on iGPU.
From OpenVINO™ 2022.3 version, the model caching feature is also supported on iGPU as preview.
This feature enables users to save and load the blob file directly. This file can be loaded directly on to the hardware device target and inferencing can be performed.
Kernel Caching on iGPU :
This feature also allows user to save kernel caching as cl_cache files for models with dynamic input shapes. These cl_cache files can be loaded directly onto the iGPU hardware device target and inferencing can be performed.
Enabling Model Caching via Runtime options using c++/python API's.
This flow can be enabled by setting the runtime config option 'use_compiled_network' to True while using the c++/python API'S. This config option acts like a switch to on and off the feature.
The blobs are saved and loaded from a directory named 'ov_compiled_blobs' relative to the executable path by default. This path however can be overridden using the runtime config option 'blob_dump_path' which is used to explicitly specify the path where you would like to dump and load the blobs (CPU, iGPU) or cl_cache(iGPU) files from when already using the use_compiled_network(model caching) setting.
Refer to Configuration Options for more information about using these runtime options.
Support for INT8 Quantized models
Int8 models are supported on CPU and GPU.
Support for Weights saved in external files
OpenVINO™ Execution Provider now supports ONNX models that store weights in external files. It is especially useful for models larger than 2GB because of protobuf limitations.
See the OpenVINO™ ONNX Support documentation.
Converting and Saving an ONNX Model to External Data: Use the ONNX API's.documentation.
Example:
import onnx
onnx_model = onnx.load("model.onnx") # Your model in memory as ModelProto
onnx.save_model(onnx_model, 'saved_model.onnx', save_as_external_data=True, all_tensors_to_one_file=True, location='data/weights_data', size_threshold=1024, convert_attribute=False)
Note:
-
In the above script, model.onnx is loaded and then gets saved into a file called 'saved_model.onnx' which won't have the weights but this new onnx model now will have the relative path to where the weights file is located. The weights file 'weights_data' will now contain the weights of the model and the weights from the original model gets saved at /data/weights_data.
-
Now, you can use this 'saved_model.onnx' file to infer using your sample. But remember, the weights file location can't be changed. The weights have to be present at /data/weights_data
-
Install the latest ONNX Python package using pip to run these ONNX Python API's successfully. .
Support for IO Buffer Optimization
To enable IO Buffer Optimization we have to set OPENCL_LIBS, OPENCL_INCS environment variables before build. For IO Buffer Optimization, the model must be fully supported on OpenVINO™ and we must provide in the remote context cl_context void pointer as C++ Configuration Option. We can provide cl::Buffer address as Input using GPU Memory Allocator for input and output.
Example:
//Set up a remote context
cl::Context _context;
.....
// Set the context through openvino options
OrtOpenVINOProviderOptions options;
options.context = (void *) _context.get() ;
.....
//Define the Memory area
Ort::MemoryInfo info_gpu("OpenVINO_GPU", OrtAllocatorType::OrtDeviceAllocator, 0, OrtMemTypeDefault);
//Create a shared buffer , fill in with data
cl::Buffer shared_buffer(_context, CL_MEM_READ_WRITE, imgSize, NULL, &err);
....
//Cast it to void*, and wrap it as device pointer for Ort::Value
void *shared_buffer_void = static_cast<void *>(&shared_buffer);
Ort::Value inputTensors = Ort::Value::CreateTensor(
info_gpu, shared_buffer_void, imgSize, inputDims.data(),
inputDims.size(), ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT);
Configuration Options
OpenVINO™ Execution Provider can be configured with certain options at runtime that control the behavior of the EP. These options can be set as key-value pairs as below:-
Python API
Key-Value pairs for config options can be set using InferenceSession API as follow:-
session = onnxruntime.InferenceSession(<path_to_model_file>, providers=['OpenVINOExecutionProvider'], provider_options=[{Key1 : Value1, Key2 : Value2, ...}])
Note that the releases from (ORT 1.10) will require explicitly setting the providers parameter if you want to use execution providers other than the default CPU provider (as opposed to the current behavior of providers getting set/registered by default based on the build flags) when instantiating InferenceSession.
C/C++ API
All the options shown below are passed to SessionOptionsAppendExecutionProvider_OpenVINO() API and populated in the struct OrtOpenVINOProviderOptions in an example shown below, for example for CPU device type:
OrtOpenVINOProviderOptions options;
options.device_type = "CPU_FP32";
options.enable_vpu_fast_compile = 0;
options.device_id = "";
options.num_of_threads = 8;
options.use_compiled_network = false;
options.blob_dump_path = "";
options.context = 0x123456ff;
options.enable_opencl_throttling = false;
SessionOptionsAppendExecutionProvider_OpenVINO(session_options, &options);
Summary of options
The following table lists all the available configuration options and the Key-Value pairs to set them:
| Key | Key type | Allowable Values | Value type | Description |
|---|---|---|---|---|
| device_type | string | CPU_FP32, CPU_FP16, GPU_FP32, GPU_FP16, MYRIAD_FP16, GPU.0_FP16, GPU.1_FP16, GPU.0_FP16, GPU.0_FP32 based on the avaialable GPUs, Any valid Hetero combination, Any valid Multi or Auto devices combination | string | Overrides the accelerator hardware type and precision with these values at runtime. If this option is not explicitly set, default hardware and precision specified during build time is used. |
| device_id | string | Any valid OpenVINO device ID | string | Selects a particular hardware device for inference. The list of valid OpenVINO device ID's available on a platform can be obtained either by Python API (onnxruntime.capi._pybind_state.get_available_openvino_device_ids()) or by OpenVINO C/C++ API. If this option is not explicitly set, an arbitrary free device will be automatically selected by OpenVINO runtime. |
| enable_vpu_fast_compile | string | True/False | boolean | This option is only available for MYRIAD_FP16 VPU devices. During initialization of the VPU device with compiled model, Fast-compile may be optionally enabled to speeds up the model's compilation to VPU device specific format. This in-turn speeds up model initialization time. However, enabling this option may slowdown inference due to some of the optimizations not being fully applied, so caution is to be exercised while enabling this option. |
| num_of_threads | string | Any unsigned positive number other than 0 | size_t | Overrides the accelerator default value of number of threads with this value at runtime. If this option is not explicitly set, default value of 8 is used during build time. |
| use_compiled_network | string | True/False | boolean | This option is only available for MYRIAD_FP16 VPU devices for both Linux and Windows and it enables save/load blob functionality. It can be used to directly import pre-compiled blobs if exists or dump a pre-compiled blob at the executable path. |
| blob_dump_path | string | Any valid string path on the hardware target | string | Explicitly specify the path where you would like to dump and load the blobs for the save/load blob feature when use_compiled_network setting is enabled . This overrides the default path. |
| context | string | OpenCL Context | void* | This option is only alvailable when OpenVINO EP is built with OpenCL flags enabled. It takes in the remote context i.e the cl_context address as a void pointer. |
| enable_opencl_throttling | string | True/False | boolean | This option enables OpenCL queue throttling for GPU devices (reduces CPU utilization when using GPU). |
| enable_dynamic_shapes | string | True/False | boolean | This option if enabled works for dynamic shaped models whose shape will be set dynamically based on the infer input image/data shape at run time in CPU. This gives best result for running multiple inferences with varied shaped images/data. |
Valid Hetero or Multi or Auto Device combinations: HETERO:<DEVICE_TYPE_1>,<DEVICE_TYPE_2>,<DEVICE_TYPE_3>... The <DEVICE_TYPE> can be any of these devices from this list ['CPU','GPU','MYRIAD']
A minimum of two DEVICE_TYPE'S should be specified for a valid HETERO or Multi-Device Build.
Example: HETERO:MYRIAD,CPU AUTO:GPU,CPU MULTI:MYRIAD,GPU,CPU
Other configuration settings
Onnxruntime Graph Optimization level
OpenVINO™ backend performs both hardware dependent as well as independent optimizations to the graph to infer it with on the target hardware with best possible performance. In most of the cases it has been observed that passing in the graph from the input model as is would lead to best possible optimizations by OpenVINO™. For this reason, it is advised to turn off high level optimizations performed by ONNX Runtime before handing the graph over to OpenVINO™ backend. This can be done using Session options as shown below:-
Python API
options = onnxruntime.SessionOptions()
options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_DISABLE_ALL
sess = onnxruntime.InferenceSession(<path_to_model_file>, options)
C/C++ API
SessionOptions::SetGraphOptimizationLevel(ORT_DISABLE_ALL);
Deprecated: Dynamic device type selection
Note: This API has been deprecated. Please use the mechanism mentioned above to set the 'device-type' option. When ONNX Runtime is built with OpenVINO™ Execution Provider, a target hardware option needs to be provided. This build time option becomes the default target harware the EP schedules inference on. However, this target may be overriden at runtime to schedule inference on a different hardware as shown below.
Note: This dynamic hardware selection is optional. The EP falls back to the build-time default selection if no dynamic hardware option value is specified.
Python API
import onnxruntime
onnxruntime.capi._pybind_state.set_openvino_device("<harware_option>")
# Create session after this
This property persists and gets applied to new sessions until it is explicity unset. To unset, assign a null string ("").
C/C++ API
Append the settings string "<hardware_option>" to the EP settings string. Example shown below for the CPU_FP32 option:
std::string settings_str;
...
settings_str.append("CPU_FP32");
Ort::ThrowOnError(OrtSessionOptionsAppendExecutionProvider_OpenVINO(sf, settings_str.c_str()));
Support Coverage
ONNX Layers supported using OpenVINO
The table below shows the ONNX layers supported and validated using OpenVINO™ Execution Provider.The below table also lists the Intel hardware support for each of the layers. CPU refers to Intel® Atom, Core, and Xeon processors. GPU refers to the Intel Integrated Graphics. Intel Discrete Graphics
| ONNX Layers | CPU | GPU |
|---|---|---|
| Abs | Yes | Yes |
| Acos | Yes | Yes |
| Acosh | Yes | Yes |
| Add | Yes | Yes |
| And | Yes | Yes |
| ArgMax | Yes | Yes |
| ArgMin | Yes | Yes |
| Asin | Yes | Yes |
| Asinh | Yes | Yes |
| Atan | Yes | Yes |
| Atanh | Yes | Yes |
| AveragePool | Yes | Yes |
| BatchNormalization | Yes | Yes |
| BitShift | Yes | No |
| Ceil | Yes | Yes |
| Celu | Yes | Yes |
| Cast | Yes | Yes |
| Clip | Yes | Yes |
| Concat | Yes | Yes |
| Constant | Yes | Yes |
| ConstantOfShape | Yes | Yes |
| Conv | Yes | Yes |
| ConvInteger | Yes | Yes |
| ConvTranspose | Yes | Yes |
| Cos | Yes | Yes |
| Cosh | Yes | Yes |
| CumSum | Yes | Yes |
| DepthToSpace | Yes | Yes |
| DequantizeLinear | Yes | Yes |
| Div | Yes | Yes |
| Dropout | Yes | Yes |
| Elu | Yes | Yes |
| Equal | Yes | Yes |
| Erf | Yes | Yes |
| Exp | Yes | Yes |
| Expand | Yes | Yes |
| EyeLike | Yes | No |
| Flatten | Yes | Yes |
| Floor | Yes | Yes |
| Gather | Yes | Yes |
| GatherElements | No | No |
| GatherND | Yes | Yes |
| Gemm | Yes | Yes |
| GlobalAveragePool | Yes | Yes |
| GlobalLpPool | Yes | Yes |
| GlobalMaxPool | Yes | Yes |
| Greater | Yes | Yes |
| GreaterOrEqual | Yes | Yes |
| GridSample | Yes | No |
| HardMax | Yes | Yes |
| HardSigmoid | Yes | Yes |
| Identity | Yes | Yes |
| If | Yes | Yes |
| ImageScaler | Yes | Yes |
| InstanceNormalization | Yes | Yes |
| LeakyRelu | Yes | Yes |
| Less | Yes | Yes |
| LessOrEqual | Yes | Yes |
| Log | Yes | Yes |
| LogSoftMax | Yes | Yes |
| Loop | Yes | Yes |
| LRN | Yes | Yes |
| LSTM | Yes | Yes |
| MatMul | Yes | Yes |
| MatMulInteger | Yes | No |
| Max | Yes | Yes |
| MaxPool | Yes | Yes |
| Mean | Yes | Yes |
| MeanVarianceNormalization | Yes | Yes |
| Min | Yes | Yes |
| Mod | Yes | Yes |
| Mul | Yes | Yes |
| Neg | Yes | Yes |
| NonMaxSuppression | Yes | Yes |
| NonZero | Yes | No |
| Not | Yes | Yes |
| OneHot | Yes | Yes |
| Or | Yes | Yes |
| Pad | Yes | Yes |
| Pow | Yes | Yes |
| PRelu | Yes | Yes |
| QuantizeLinear | Yes | Yes |
| QLinearMatMul | Yes | No |
| Range | Yes | Yes |
| Reciprocal | Yes | Yes |
| ReduceL1 | Yes | Yes |
| ReduceL2 | Yes | Yes |
| ReduceLogSum | Yes | Yes |
| ReduceLogSumExp | Yes | Yes |
| ReduceMax | Yes | Yes |
| ReduceMean | Yes | Yes |
| ReduceMin | Yes | Yes |
| ReduceProd | Yes | Yes |
| ReduceSum | Yes | Yes |
| ReduceSumSquare | Yes | Yes |
| Relu | Yes | Yes |
| Reshape | Yes | Yes |
| Resize | Yes | Yes |
| ReverseSequence | Yes | Yes |
| RoiAlign | Yes | Yes |
| Round | Yes | Yes |
| Scatter | Yes | Yes |
| ScatterElements | Yes | Yes |
| ScatterND | Yes | Yes |
| Selu | Yes | Yes |
| Shape | Yes | Yes |
| Shrink | Yes | Yes |
| Sigmoid | Yes | Yes |
| Sign | Yes | Yes |
| Sin | Yes | Yes |
| Sinh | Yes | No |
| SinFloat | No | No |
| Size | Yes | Yes |
| Slice | Yes | Yes |
| Softmax | Yes | Yes |
| Softplus | Yes | Yes |
| Softsign | Yes | Yes |
| SpaceToDepth | Yes | Yes |
| Split | Yes | Yes |
| Sqrt | Yes | Yes |
| Squeeze | Yes | Yes |
| Sub | Yes | Yes |
| Sum | Yes | Yes |
| Softsign | Yes | No |
| Tan | Yes | Yes |
| Tanh | Yes | Yes |
| ThresholdedRelu | Yes | Yes |
| Tile | Yes | Yes |
| TopK | Yes | Yes |
| Transpose | Yes | Yes |
| Unsqueeze | Yes | Yes |
| Upsample | Yes | Yes |
| Where | Yes | Yes |
| Xor | Yes | Yes |
Topology Support
Below topologies from ONNX open model zoo are fully supported on OpenVINO™ Execution Provider and many more are supported through sub-graph partitioning
Image Classification Networks
| MODEL NAME | CPU | GPU |
|---|---|---|
| bvlc_alexnet | Yes | Yes |
| bvlc_googlenet | Yes | Yes |
| bvlc_reference_caffenet | Yes | Yes |
| bvlc_reference_rcnn_ilsvrc13 | Yes | Yes |
| emotion ferplus | Yes | Yes |
| densenet121 | Yes | Yes |
| inception_v1 | Yes | Yes |
| inception_v2 | Yes | Yes |
| mobilenetv2 | Yes | Yes |
| resnet18v1 | Yes | Yes |
| resnet34v1 | Yes | Yes |
| resnet101v1 | Yes | Yes |
| resnet152v1 | Yes | Yes |
| resnet18v2 | Yes | Yes |
| resnet34v2 | Yes | Yes |
| resnet101v2 | Yes | Yes |
| resnet152v2 | Yes | Yes |
| resnet50 | Yes | Yes |
| resnet50v2 | Yes | Yes |
| shufflenet | Yes | Yes |
| squeezenet1.1 | Yes | Yes |
| vgg19 | Yes | Yes |
| vgg16 | Yes | Yes |
| zfnet512 | Yes | Yes |
| mxnet_arcface | No | Yes |
Image Recognition Networks
| MODEL NAME | CPU | GPU |
|---|---|---|
| mnist | Yes | Yes |
Object Detection Networks
| MODEL NAME | CPU | GPU |
|---|---|---|
| tiny_yolov2 | Yes | Yes |
| yolov3 | Yes | Yes |
| tiny_yolov3 | Yes | Yes |
| mask_rcnn | Yes | Yes |
| faster_rcnn | Yes | Yes |
| yolov4 | Yes | Yes |
| yolov5 | Yes | Yes |
| yolov7 | Yes | Yes |
| tiny_yolov7 | Yes | Yes |
Image Manipulation Networks
| MODEL NAME | CPU | GPU |
|---|---|---|
| mosaic | Yes | Yes |
| candy | Yes | Yes |
| cgan | Yes | Yes |
| rain_princess | Yes | Yes |
| pointilism | Yes | Yes |
| udnie | Yes | Yes |
Natural Language Processing Networks
| MODEL NAME | CPU | GPU |
|---|---|---|
| bert-squad | Yes | Yes |
| bert-base-cased | Yes | Yes |
| bert-base-chinese | Yes | Yes |
| bert-base-japanese-char | Yes | Yes |
| bert-base-multilingual-cased | Yes | Yes |
| bert-base-uncased | Yes | Yes |
| distilbert-base-cased | Yes | Yes |
| distilbert-base-multilingual-cased | Yes | Yes |
| distilbert-base-uncased | Yes | Yes |
| distilbert-base-uncased-finetuned-sst-2-english | Yes | Yes |
| gpt2 | Yes | Yes |
| roberta-base | Yes | Yes |
| roberta-base-squad2 | Yes | Yes |
| t5-base | Yes | Yes |
| twitter-roberta-base-sentiment | Yes | Yes |
| xlm-roberta-base | Yes | Yes |
Note: We have added support for INT8 models, quantized with Neural Network Compression Framework (NNCF). To know more about NNCF refer here.
OpenVINO™ Execution Provider Samples Tutorials
In order to showcase what you can do with the OpenVINO™ Execution Provider for ONNX Runtime, we have created a few samples that shows how you can get that performance boost you’re looking for with just one additional line of code.
Python API
Object detection with tinyYOLOv2 in Python
Object detection with YOLOv4 in Python
C/C++ API
Image classification with Squeezenet in CPP
Csharp API
Object detection with YOLOv3 in C#
Blogs/Tutorials
Overview of OpenVINO Execution Provider for ONNX Runtime
Tutorial on how to use OpenVINO™ Execution Provider for ONNX Runtime Docker Containers
Tutorial on how to use OpenVINO™ Execution Provider for ONNX Runtime python wheel packages
Accelerate inference for PyTorch models with OpenVINO Execution Provider (Preview)
ONNX Runtime for PyTorch is now extended to support PyTorch model inference using ONNX Runtime. It is available via the torch-ort-infer python package. This preview package enables OpenVINO™ Execution Provider for ONNX Runtime by default for accelerating inference on various Intel® CPUs, Intel® integrated GPUs, and Intel® Movidius™ Vision Processing Units - referred to as VPU. For more details, see torch-ort-infer.