OpenVINO Plugin usually represents a wrapper around a backend. Backends can be:

  • OpenCL-like backend (e.g. clDNN library) for GPU devices.

  • oneDNN backend for Intel CPU devices.


The responsibility of OpenVINO Plugin:

  • Initializes a backend and throw exception in Engine constructor if backend cannot be initialized.

  • Provides information about devices enabled by a particular backend, e.g. how many devices, their properties and so on.

  • Loads or imports compiled model objects.

In addition to the OpenVINO Public API, the OpenVINO provides the Plugin API, which is a set of functions and helper classes that simplify new plugin development:

  • header files in the src/inference/dev_api/openvino directory

  • implementations in the src/inference/src/dev/ directory

  • symbols in the OpenVINO shared library

To build an OpenVINO plugin with the Plugin API, see the OpenVINO Plugin Building guide.

Plugin Class

OpenVINO Plugin API provides the helper ov::IPlugin class recommended to use as a base class for a plugin. Based on that, declaration of a plugin class can look as follows:

namespace ov {
namespace template_plugin {

class Plugin : public ov::IPlugin {

    std::shared_ptr<ov::ICompiledModel> compile_model(const std::shared_ptr<const ov::Model>& model,
                                                      const ov::AnyMap& properties) const override;

    std::shared_ptr<ov::ICompiledModel> compile_model(const std::shared_ptr<const ov::Model>& model,
                                                      const ov::AnyMap& properties,
                                                      const ov::SoPtr<ov::IRemoteContext>& context) const override;

    void set_property(const ov::AnyMap& properties) override;

    ov::Any get_property(const std::string& name, const ov::AnyMap& arguments) const override;

    ov::SoPtr<ov::IRemoteContext> create_context(const ov::AnyMap& remote_properties) const override;

    ov::SoPtr<ov::IRemoteContext> get_default_context(const ov::AnyMap& remote_properties) const override;

    std::shared_ptr<ov::ICompiledModel> import_model(std::istream& model, const ov::AnyMap& properties) const override;

    std::shared_ptr<ov::ICompiledModel> import_model(std::istream& model,
                                                     const ov::SoPtr<ov::IRemoteContext>& context,
                                                     const ov::AnyMap& properties) const override;

    ov::SupportedOpsMap query_model(const std::shared_ptr<const ov::Model>& model,
                                    const ov::AnyMap& properties) const override;

    friend class CompiledModel;
    friend class InferRequest;

    std::shared_ptr<ov::runtime::Backend> m_backend;
    Configuration m_cfg;
    std::shared_ptr<ov::threading::ITaskExecutor> m_waitExecutor;

}  // namespace template_plugin
}  // namespace ov

Class Fields

The provided plugin class also has several fields:

  • m_backend - a backend engine that is used to perform actual computations for model inference. For Template plugin ov::runtime::Backend is used which performs computations using OpenVINO™ reference implementations.

  • m_waitExecutor - a task executor that waits for a response from a device about device tasks completion.

  • m_cfg of type Configuration:

struct Configuration {
    Configuration(const Configuration&) = default;
    Configuration(Configuration&&) = default;
    Configuration& operator=(const Configuration&) = default;
    Configuration& operator=(Configuration&&) = default;

    explicit Configuration(const ov::AnyMap& config,
                           const Configuration& defaultCfg = {},
                           const bool throwOnUnsupported = true);

    ov::Any Get(const std::string& name) const;

    // Plugin configuration parameters

    int device_id = 0;
    bool perf_count = false;
    ov::threading::IStreamsExecutor::Config streams_executor_config;
    ov::hint::PerformanceMode performance_mode = ov::hint::PerformanceMode::LATENCY;
    uint32_t num_requests = 1;
    bool disable_transformations = false;
    bool exclusive_async_requests = false;

    // unused
    ov::element::Type inference_precision = ov::element::undefined;
    ov::hint::ExecutionMode execution_mode = ov::hint::ExecutionMode::ACCURACY;
    ov::log::Level log_level = ov::log::Level::NO;

    ov::hint::Priority model_priority = ov::hint::Priority::DEFAULT;

As an example, a plugin configuration has three value parameters:

  • device_id - particular device ID to work with. Applicable if a plugin supports more than one Template device. In this case, some plugin methods, like set_property, query_model, and compile_model, must support the ov::device::id property.

  • perf_counts - boolean value to identify whether to collect performance counters during Inference Request execution.

  • streams_executor_config - configuration of ov::threading::IStreamsExecutor to handle settings of multi-threaded context.

  • performance_mode - configuration of ov::hint::PerformanceMode to set the performance mode.

  • disable_transformations - allows to disable transformations which are applied in the process of model compilation.

  • exclusive_async_requests - allows to use exclusive task executor for asynchronous infer requests.

Plugin Constructor

A plugin constructor must contain code that checks the ability to work with a device of the Template type. For example, if some drivers are required, the code must check driver availability. If a driver is not available (for example, OpenCL runtime is not installed in case of a GPU device or there is an improper version of a driver is on a host machine), an exception must be thrown from a plugin constructor.

A plugin must define a device name enabled via the set_device_name() method of a base class:

ov::template_plugin::Plugin::Plugin() {
    // TODO: fill with actual device name, backend engine

    // create backend which performs inference using openvino reference implementations
    m_backend = ov::runtime::Backend::create();

    // create default stream executor with a given name
    m_waitExecutor = get_executor_manager()->get_idle_cpu_streams_executor({wait_executor_name});

Plugin Destructor

A plugin destructor must stop all plugins activities, and clean all allocated resources.

ov::template_plugin::Plugin::~Plugin() {
    // Plugin should remove executors from executor cache to avoid threads number growth in the whole application


The plugin should implement two compile_model() methods: the first one compiles model without remote context, the second one with remote context if plugin supports.

This is the most important function of the Plugin class is to create an instance of compiled CompiledModel, which holds a backend-dependent compiled model in an internal representation:

std::shared_ptr<ov::ICompiledModel> ov::template_plugin::Plugin::compile_model(
    const std::shared_ptr<const ov::Model>& model,
    const ov::AnyMap& properties) const {
    return compile_model(model, properties, {});
std::shared_ptr<ov::ICompiledModel> ov::template_plugin::Plugin::compile_model(
    const std::shared_ptr<const ov::Model>& model,
    const ov::AnyMap& properties,
    const ov::SoPtr<ov::IRemoteContext>& context) const {
    OV_ITT_SCOPED_TASK(itt::domains::TemplatePlugin, "Plugin::compile_model");

    auto fullConfig = Configuration{properties, m_cfg};
    auto streamsExecutorConfig =
    streamsExecutorConfig._name = stream_executor_name;
    auto compiled_model = std::make_shared<CompiledModel>(
            ? get_executor_manager()->get_executor(template_exclusive_executor)
            : get_executor_manager()->get_idle_cpu_streams_executor(streamsExecutorConfig),
    return compiled_model;

Before a creation of an CompiledModel instance via a constructor, a plugin may check if a provided ov::Model object is supported by a device if it is needed.

Actual model compilation is done in the CompiledModel constructor. Refer to the CompiledModel Implementation Guide for details.


Actual configuration map used in CompiledModel is constructed as a base plugin configuration set via Plugin::set_property, where some values are overwritten with config passed to Plugin::compile_model. Therefore, the config of Plugin::compile_model has a higher priority.


The function accepts a const shared pointer to ov::Model object and applies common and device-specific transformations on a copied model to make it more friendly to hardware operations. For details how to write custom device-specific transformation, refer to Writing OpenVINO™ transformations guide. See detailed topics about model representation:

void transform_model(const std::shared_ptr<ov::Model>& model) {
    // Perform common optimizations and device-specific transformations
    ov::pass::Manager passManager;
    // Example: register CommonOptimizations transformation from transformations library
    // Disable some transformations
    // This transformation changes output name
    // Register any other transformations
    // ..

    const auto& pass_config = passManager.get_pass_config();

    // Allow FP16 Converts to be folded and FP16 constants to be upgraded to FP32 data type

    // After `run_passes`, we have the transformed function, where operations match device operations,
    // and we can create device backend-dependent graph


After all these transformations, an ov::Model object contains operations which can be perfectly mapped to backend kernels. E.g. if backend has kernel computing A + B operations at once, the transform_model function should contain a pass which fuses operations A and B into a single custom operation A + B which fits backend kernels set.


Use the method with the HETERO mode, which allows to distribute model execution between different devices based on the ov::Node::get_rt_info() map, which can contain the affinity key. The query_model method analyzes operations of provided model and returns a list of supported operations via the ov::SupportedOpsMap structure. The query_model firstly applies transform_model passes to input ov::Model argument. After this, the transformed model in ideal case contains only operations are 1:1 mapped to kernels in computational backend. In this case, it’s very easy to analyze which operations is supposed (m_backend has a kernel for such operation or extensions for the operation is provided) and not supported (kernel is missed in m_backend):

  1. Store original names of all operations in input ov::Model.

  2. Apply transform_model passes. Note, the names of operations in a transformed model can be different and we need to restore the mapping in the steps below.

  3. Construct supported map which contains names of original operations. Note that since the inference is performed using OpenVINO™ reference backend, the decision whether the operation is supported or not depends on whether the latest OpenVINO opset contains such operation.

  4. ov.SupportedOpsMap contains only operations which are fully supported by m_backend.

ov::SupportedOpsMap ov::template_plugin::Plugin::query_model(const std::shared_ptr<const ov::Model>& model,
                                                             const ov::AnyMap& properties) const {
    OV_ITT_SCOPED_TASK(itt::domains::TemplatePlugin, "Plugin::query_model");

    Configuration fullConfig{properties, m_cfg, false};

    OPENVINO_ASSERT(model, "OpenVINO Model is empty!");

    auto supported = ov::get_supported_nodes(
        [&](std::shared_ptr<ov::Model>& model) {
            // skip transformations in case of user config
            if (fullConfig.disable_transformations)
            // 1. It is needed to apply all transformations as it is done in compile_model
        [&](std::shared_ptr<ov::Node> node) {
            // 2. Сheck whether node is supported
            ov::OpSet op_super_set;
#define _OPENVINO_OP_REG(NAME, NAMESPACE) op_super_set.insert<NAMESPACE::NAME>();
        // clang-format off
#include "openvino/opsets/opset1_tbl.hpp"
#include "openvino/opsets/opset2_tbl.hpp"
#include "openvino/opsets/opset3_tbl.hpp"
#include "openvino/opsets/opset4_tbl.hpp"
#include "openvino/opsets/opset5_tbl.hpp"
#include "openvino/opsets/opset6_tbl.hpp"
#include "openvino/opsets/opset7_tbl.hpp"
#include "openvino/opsets/opset8_tbl.hpp"
#include "openvino/opsets/opset9_tbl.hpp"
#include "openvino/opsets/opset10_tbl.hpp"
#include "openvino/opsets/opset11_tbl.hpp"
#include "openvino/opsets/opset12_tbl.hpp"
#include "openvino/opsets/opset13_tbl.hpp"
        // clang-format on
            return op_super_set.contains_type(node->get_type_info());

    // 3. Produce the result
    ov::SupportedOpsMap res;
    for (auto&& layerName : supported) {
        res.emplace(layerName, get_device_name() + "." + std::to_string(m_cfg.device_id));

    return res;


Sets new values for plugin property keys:

void ov::template_plugin::Plugin::set_property(const ov::AnyMap& properties) {
    m_cfg = Configuration{properties, m_cfg};

In the snippet above, the Configuration class overrides previous configuration values with the new ones. All these values are used during backend specific model compilation and execution of inference requests.


The function must throw an exception if it receives an unsupported configuration key.


Returns a current value for a specified property key:

ov::Any ov::template_plugin::Plugin::get_property(const std::string& name, const ov::AnyMap& arguments) const {
    const auto& default_ro_properties = []() {
        std::vector<ov::PropertyName> ro_properties{ov::available_devices,
        return ro_properties;
    const auto& default_rw_properties = []() {
        std::vector<ov::PropertyName> rw_properties{ov::device::id,
        return rw_properties;
    if (ov::supported_properties == name) {
        auto ro_properties = default_ro_properties();
        auto rw_properties = default_rw_properties();

        std::vector<ov::PropertyName> supported_properties;
        supported_properties.reserve(ro_properties.size() + rw_properties.size());
        supported_properties.insert(supported_properties.end(), ro_properties.begin(), ro_properties.end());
        supported_properties.insert(supported_properties.end(), rw_properties.begin(), rw_properties.end());
        return decltype(ov::supported_properties)::value_type(supported_properties);
    } else if (ov::internal::supported_properties == name) {
        return decltype(ov::internal::supported_properties)::value_type{
            ov::PropertyName{, ov::PropertyMutability::RO},
            ov::PropertyName{, ov::PropertyMutability::RW}};
    } else if (ov::available_devices == name) {
        // TODO: fill list of available devices
        std::vector<std::string> available_devices = {""};
        return decltype(ov::available_devices)::value_type(available_devices);
    } else if (ov::device::full_name == name) {
        std::string device_name = "Template Device Full Name";
        return decltype(ov::device::full_name)::value_type(device_name);
    } else if (ov::device::architecture == name) {
        // TODO: return device architecture for device specified by DEVICE_ID config
        std::string arch = get_device_name();
        return decltype(ov::device::architecture)::value_type(arch);
    } else if (ov::device::type == name) {
        return decltype(ov::device::type)::value_type(ov::device::Type::INTEGRATED);
    } else if (ov::internal::caching_properties == name) {
        std::vector<ov::PropertyName> caching_properties = {ov::device::architecture};
        return decltype(ov::internal::caching_properties)::value_type(caching_properties);
    } else if (ov::device::capabilities == name) {
        // TODO: fill actual list of supported capabilities: e.g. Template device supports only FP32 and EXPORT_IMPORT
        std::vector<std::string> capabilities = {ov::device::capability::FP32, ov::device::capability::EXPORT_IMPORT};
        return decltype(ov::device::capabilities)::value_type(capabilities);
    } else if (ov::execution_devices == name) {
        std::string dev = get_device_name();
        return decltype(ov::execution_devices)::value_type{dev};
    } else if (ov::range_for_async_infer_requests == name) {
        // TODO: fill with actual values
        using uint = unsigned int;
        return decltype(ov::range_for_async_infer_requests)::value_type(std::make_tuple(uint{1}, uint{1}, uint{1}));
    } else {
        return m_cfg.Get(name);

The function is implemented with the Configuration::Get method, which wraps an actual configuration key value to the ov::Any and returns it.


The function must throw an exception if it receives an unsupported configuration key.


The importing of compiled model mechanism allows to import a previously exported backend specific model and wrap it using an CompiledModel object. This functionality is useful if backend specific model compilation takes significant time and/or cannot be done on a target host device due to other reasons.

During export of backend specific model using CompiledModel::export_model, a plugin may export any type of information it needs to import a compiled model properly and check its correctness. For example, the export information may include:

  • Compilation options (state of Plugin::m_cfg structure).

  • Information about a plugin and a device type to check this information later during the import and throw an exception if the model stream contains wrong data. For example, if devices have different capabilities and a model compiled for a particular device cannot be used for another, such type of information must be stored and checked during the import.

  • Compiled backend specific model itself.

std::shared_ptr<ov::ICompiledModel> ov::template_plugin::Plugin::import_model(std::istream& model,
                                                                              const ov::AnyMap& properties) const {
    return import_model(model, {}, properties);
std::shared_ptr<ov::ICompiledModel> ov::template_plugin::Plugin::import_model(
    std::istream& model,
    const ov::SoPtr<ov::IRemoteContext>& context,
    const ov::AnyMap& properties) const {
    OV_ITT_SCOPED_TASK(itt::domains::TemplatePlugin, "Plugin::import_model");

    auto fullConfig = Configuration{properties, m_cfg};
    // read XML content
    std::string xmlString;
    std::uint64_t dataSize = 0;<char*>(&dataSize), sizeof(dataSize));
    xmlString.resize(dataSize);<char*>(xmlString.c_str()), dataSize);

    // read blob content
    ov::Tensor weights;<char*>(&dataSize), sizeof(dataSize));
    if (0 != dataSize) {
        weights = ov::Tensor(ov::element::from<char>(), ov::Shape{static_cast<ov::Shape::size_type>(dataSize)});<char>(), dataSize);

    auto ov_model = get_core()->read_model(xmlString, weights);
    auto streamsExecutorConfig =
    streamsExecutorConfig._name = stream_executor_name;
    auto compiled_model =
    return compiled_model;


The Plugin should implement Plugin::create_context() method which returns ov::RemoteContext in case if plugin supports remote context, in other case the plugin can throw an exception that this method is not implemented.

ov::SoPtr<ov::IRemoteContext> ov::template_plugin::Plugin::create_context(const ov::AnyMap& remote_properties) const {
    return std::make_shared<ov::template_plugin::RemoteContext>();


Plugin::get_default_context() also needed in case if plugin supports remote context, if the plugin doesn’t support it, this method can throw an exception that functionality is not implemented.

ov::SoPtr<ov::IRemoteContext> ov::template_plugin::Plugin::get_default_context(
    const ov::AnyMap& remote_properties) const {
    return std::make_shared<ov::template_plugin::RemoteContext>();

Create Instance of Plugin Class

OpenVINO plugin library must export only one function creating a plugin instance using OV_DEFINE_PLUGIN_CREATE_FUNCTION macro:

static const ov::Version version = {CI_BUILD_NUMBER, "openvino_template_plugin"};
OV_DEFINE_PLUGIN_CREATE_FUNCTION(ov::template_plugin::Plugin, version)

Next step in a plugin library implementation is the CompiledModel class.