[NPU] Move network_metadata from AL folder to Common

src/plugins/intel_npu/src/backend/include/zero_infer_request.hpp

@@ -6,8 +6,8 @@
 
 #include "intel_npu/common/icompiled_model.hpp"
 #include "intel_npu/common/igraph.hpp"
+#include "intel_npu/common/network_metadata.hpp"
 #include "intel_npu/common/npu.hpp"
-#include "intel_npu/network_metadata.hpp"
 #include "intel_npu/utils/logger/logger.hpp"
 #include "intel_npu/utils/zero/zero_tensor.hpp"
 #include "zero_pipeline.hpp"

src/plugins/intel_npu/src/common/include/intel_npu/common/igraph.hpp

@@ -10,7 +10,7 @@
 #include <vector>
 
 #include "intel_npu/common/filtered_config.hpp"
-#include "intel_npu/network_metadata.hpp"
+#include "intel_npu/common/network_metadata.hpp"
 #include "intel_npu/utils/zero/zero_wrappers.hpp"
 #include "openvino/runtime/itensor.hpp"
 #include "openvino/runtime/profiling_info.hpp"

src/plugins/intel_npu/src/al/include/intel_npu/network_metadata.hpp → src/plugins/intel_npu/src/common/include/intel_npu/common/network_metadata.hpp

@@ -158,26 +158,4 @@ struct NetworkMetadata final {
     void bindRelatedDescriptors();
 };
 
-/**
- * @struct NetworkDescription
- * @brief The object returned by the compiler
- * to provide such information about a network as description of inputs and outputs,
- * name and compiled network in a format executable by device
- */
-struct NetworkDescription final {
-    NetworkDescription(ov::Tensor&& compiledNetWorkTensor, NetworkMetadata&& metadata)
-        : metadata(std::move(metadata)),
-          compiledNetworkTensor(std::move(compiledNetWorkTensor)) {}
-    // Force move semantics to prevent blob copies
-    NetworkDescription(const NetworkDescription&) = delete;
-    NetworkDescription(NetworkDescription&&) = default;
-    NetworkDescription& operator=(const NetworkDescription&) = delete;
-    NetworkDescription& operator=(NetworkDescription&&) = default;
-    ~NetworkDescription() = default;
-
-    NetworkMetadata metadata;
-
-    ov::Tensor compiledNetworkTensor;
-};
-
 }  // namespace intel_npu

src/plugins/intel_npu/src/al/src/network_metadata.cpp → src/plugins/intel_npu/src/common/src/network_metadata.cpp

@@ -2,7 +2,7 @@
 // SPDX-License-Identifier: Apache-2.0
 //
 
-#include "intel_npu/network_metadata.hpp"
+#include "intel_npu/common/network_metadata.hpp"
 
 namespace intel_npu {
 

src/plugins/intel_npu/src/compiler_adapter/include/compiler_impl.hpp

@@ -9,7 +9,7 @@
 
 #include "compiler.h"
 #include "intel_npu/common/filtered_config.hpp"
-#include "intel_npu/network_metadata.hpp"
+#include "intel_npu/common/network_metadata.hpp"
 #include "openvino/core/except.hpp"
 
 namespace intel_npu {
@@ -26,18 +26,18 @@ class VCLCompilerImpl final : public std::enable_shared_from_this<VCLCompilerImp
      * @param model a shared pointer to the OpenVINO model to be compiled
      * @param config a reference to NPUConfig containing plugin config options
      *        including config options related to compilation
-     * @return a shared pointer on an object implementing NetworkDescription interface
+     * @return an ov::Tensor object containing the blob of the compiled model
      */
-    NetworkDescription compile(const std::shared_ptr<const ov::Model>& model, const FilteredConfig& config) const;
+    ov::Tensor compile(const std::shared_ptr<const ov::Model>& model, const FilteredConfig& config) const;
 
     /**
      * @brief Compiles the model, weights separation enabled. All init schedules along with the main one are compiled in
      * the same scope.
-     * @return A "NetworkDescription" object for each init schedule, followed by another one corresponding to the main
+     * @return An ov::Tensor object for each init schedule, followed by another one corresponding to the main
      * part.
      */
-    std::vector<std::shared_ptr<NetworkDescription>> compileWsOneShot(const std::shared_ptr<ov::Model>& model,
-                                                                      const FilteredConfig& config) const;
+    std::vector<ov::Tensor> compileWsOneShot(const std::shared_ptr<ov::Model>& model,
+                                             const FilteredConfig& config) const;
     /**
      * @brief Sequential compilation of Init(s) and Main
      *
@@ -52,9 +52,9 @@ class VCLCompilerImpl final : public std::enable_shared_from_this<VCLCompilerImp
      * Compiler should somehow understand wich Init(or Main) to return
      * Plugin does not know total numbers of Init schedules
      */
-    NetworkDescription compileWsIterative(const std::shared_ptr<ov::Model>& model,
-                                          const FilteredConfig& config,
-                                          size_t callNumber) const;
+    ov::Tensor compileWsIterative(const std::shared_ptr<ov::Model>& model,
+                                  const FilteredConfig& config,
+                                  size_t callNumber) const;
     /**
      * @brief Returns information about supported layers of the network passed
      * @param model The model to be queried
@@ -102,9 +102,9 @@ class VCLCompilerImpl final : public std::enable_shared_from_this<VCLCompilerImp
      * "WeightlessCacheAttribute" may be stored within the serialized model if requested.
      * @note Storing the "WeightlessCacheAttribute" is necessary if the "weights separation" flow is being used.
      */
-    NetworkDescription compile(const std::shared_ptr<const ov::Model>& model,
-                               const FilteredConfig& config,
-                               const bool storeWeightlessCacheAttributeFlag) const;
+    ov::Tensor compile(const std::shared_ptr<const ov::Model>& model,
+                       const FilteredConfig& config,
+                       const bool storeWeightlessCacheAttributeFlag) const;
 
     vcl_log_handle_t _logHandle = nullptr;
     vcl_compiler_handle_t _compilerHandle = nullptr;

src/plugins/intel_npu/src/compiler_adapter/include/dynamic_graph.hpp

@@ -7,7 +7,7 @@
 #include <ze_graph_ext.h>
 
 #include "intel_npu/common/idynamic_graph.hpp"
-#include "intel_npu/network_metadata.hpp"
+#include "intel_npu/common/network_metadata.hpp"
 #include "intel_npu/utils/zero/zero_init.hpp"
 #include "npu_vm_runtime_api.hpp"
 #include "openvino/runtime/so_ptr.hpp"

src/plugins/intel_npu/src/compiler_adapter/include/weightless_graph.hpp

@@ -111,7 +111,6 @@ class WeightlessGraph final : public Graph {
     std::vector<std::unique_ptr<CommandList>> _initsCommandLists;
     std::vector<std::unique_ptr<Fence>> _initsFences;
     std::shared_ptr<CommandQueue> _initsCommandQueue;
-    uint32_t _initsCommandQueueGroupOrdinal = 0;
 
     /**
      * @brief Tensors holding the L0 buffers corresponding to the inputs of the main schedule.

src/plugins/intel_npu/src/compiler_adapter/include/weightless_utils.hpp

@@ -4,7 +4,7 @@
 
 #pragma once
 
-#include "intel_npu/network_metadata.hpp"
+#include "intel_npu/common/network_metadata.hpp"
 
 namespace intel_npu {
 

src/plugins/intel_npu/src/compiler_adapter/include/ze_graph_ext_wrappers.hpp

@@ -13,7 +13,7 @@
 #include <unordered_set>
 #include <vector>
 
-#include "intel_npu/network_metadata.hpp"
+#include "intel_npu/common/network_metadata.hpp"
 #include "intel_npu/utils/logger/logger.hpp"
 #include "intel_npu/utils/zero/zero_init.hpp"
 #include "model_serializer.hpp"

src/plugins/intel_npu/src/compiler_adapter/src/compiler_impl.cpp

@@ -372,14 +372,13 @@ std::shared_ptr<void> VCLCompilerImpl::getLinkedLibrary() const {
     return VCLApi::getInstance()->getLibrary();
 }
 
-NetworkDescription VCLCompilerImpl::compile(const std::shared_ptr<const ov::Model>& model,
-                                            const FilteredConfig& config) const {
+ov::Tensor VCLCompilerImpl::compile(const std::shared_ptr<const ov::Model>& model, const FilteredConfig& config) const {
     return compile(model, config, false);
 }
 
-NetworkDescription VCLCompilerImpl::compile(const std::shared_ptr<const ov::Model>& model,
-                                            const FilteredConfig& config,
-                                            const bool storeWeightlessCacheAttributeFlag) const {
+ov::Tensor VCLCompilerImpl::compile(const std::shared_ptr<const ov::Model>& model,
+                                    const FilteredConfig& config,
+                                    const bool storeWeightlessCacheAttributeFlag) const {
     _logger.debug("compile start");
 
     /// Check the linked vcl version whether supported in plugin
@@ -454,22 +453,17 @@ NetworkDescription VCLCompilerImpl::compile(const std::shared_ptr<const ov::Mode
                       allocator.m_size,
                       static_cast<void*>(allocator.m_allocated));
 
-        // Use empty metadata as VCL does not support metadata extraction
-        NetworkMetadata metadata;
-
         _logger.debug("compile end, blob size:%d", allocator.m_size);
-        return NetworkDescription(make_tensor_from_aligned_addr(allocator.m_allocated, allocator.m_size),
-                                  std::move(metadata));
+        return make_tensor_from_aligned_addr(allocator.m_allocated, allocator.m_size);
     } else {
         OPENVINO_THROW("Not supported VCL version: %d.%d, please use VCL 6.1 or later",
                        _vclVersion.major,
                        _vclVersion.minor);
     }
 }
 
-std::vector<std::shared_ptr<NetworkDescription>> VCLCompilerImpl::compileWsOneShot(
-    const std::shared_ptr<ov::Model>& model,
-    const FilteredConfig& config) const {
+std::vector<ov::Tensor> VCLCompilerImpl::compileWsOneShot(const std::shared_ptr<ov::Model>& model,
+                                                          const FilteredConfig& config) const {
     _logger.debug("compileWsOneShot start");
 
     /// Check the linked vcl version whether supported in plugin
@@ -529,20 +523,16 @@ std::vector<std::shared_ptr<NetworkDescription>> VCLCompilerImpl::compileWsOneSh
         OPENVINO_THROW("Failed to create VCL executable, blobCount is zero");
     }
 
-    std::vector<std::shared_ptr<NetworkDescription>> networkDescrs;
+    std::vector<ov::Tensor> initMainTensors;
     for (auto& blob : allocator.m_info) {
-        // Use empty metadata as VCL does not support metadata extraction
-        NetworkMetadata metadata;
-        networkDescrs.emplace_back(
-            std::make_shared<NetworkDescription>(make_tensor_from_aligned_addr(blob.first, blob.second),
-                                                 std::move(metadata)));
+        initMainTensors.emplace_back(make_tensor_from_aligned_addr(blob.first, blob.second));
     }
-    return networkDescrs;
+    return initMainTensors;
 }
 
-NetworkDescription VCLCompilerImpl::compileWsIterative(const std::shared_ptr<ov::Model>& model,
-                                                       const FilteredConfig& config,
-                                                       size_t callNumber) const {
+ov::Tensor VCLCompilerImpl::compileWsIterative(const std::shared_ptr<ov::Model>& model,
+                                               const FilteredConfig& config,
+                                               size_t callNumber) const {
     _logger.debug("compileWsIterative start");
     FilteredConfig updatedConfig = config;
     updatedConfig.update({{ov::intel_npu::ws_compile_call_number.name(), std::to_string(callNumber)}});

src/plugins/intel_npu/src/compiler_adapter/src/plugin_compiler_adapter.cpp

@@ -64,12 +64,9 @@ std::shared_ptr<IGraph> PluginCompilerAdapter::compile(const std::shared_ptr<con
     OV_ITT_TASK_CHAIN(COMPILE_BLOB, itt::domains::NPUPlugin, "PluginCompilerAdapter", "compile");
 
     _logger.debug("compile start");
-    auto networkDesc = _compiler->compile(model, config);
+    auto tensor = _compiler->compile(model, config);
     _logger.debug("compile end");
 
-    ov::Tensor tensor;
-    tensor = std::move(networkDesc.compiledNetworkTensor);
-
     if (config.get<COMPILATION_MODE>() == "HostCompile") {
         // metadata will be obtained in initialze() of DynamicGraph
         _logger.debug("Use dynamicGraph to hold blob for HostCompile mode!");
@@ -133,20 +130,17 @@ std::shared_ptr<IGraph> PluginCompilerAdapter::compileWS(std::shared_ptr<ov::Mod
 
     switch (localConfig.get<SEPARATE_WEIGHTS_VERSION>()) {
     case ov::intel_npu::WSVersion::ONE_SHOT: {
-        std::vector<std::shared_ptr<NetworkDescription>> initMainNetworkDescriptions =
-            _compiler->compileWsOneShot(model, localConfig);
+        std::vector<ov::Tensor> initMainTensors = _compiler->compileWsOneShot(model, localConfig);
 
-        std::shared_ptr<NetworkDescription> mainNetworkDescription = initMainNetworkDescriptions.back();
-        initMainNetworkDescriptions.pop_back();
-        if (initMainNetworkDescriptions.empty()) {
+        auto tensorMain = initMainTensors.back();
+        initMainTensors.pop_back();
+        if (initMainTensors.empty()) {
             _logger.warning("NPU compiler did not produce any init schedules. "
                             "This likely means that the compiled model blob has weights inside even "
                             "though weightless compilation was requested.");
         }
 
-        std::vector<std::shared_ptr<NetworkDescription>> initNetworkDescriptions =
-            std::move(initMainNetworkDescriptions);
-        tensorMain = std::move(mainNetworkDescription->compiledNetworkTensor);
+        tensorsInits = std::move(initMainTensors);
 
         if (_zeGraphExt) {
             // Depending on the config, we may get an error when trying to
@@ -164,13 +158,9 @@ std::shared_ptr<IGraph> PluginCompilerAdapter::compileWS(std::shared_ptr<ov::Mod
                 "No driver is found, zeGraphExt is nullptr, so metadata is empty. Only exports are available");
         }
 
-        initGraphDescriptors.reserve(initNetworkDescriptions.size());
-        tensorsInits.reserve(initNetworkDescriptions.size());
-        initNetworkMetadata.reserve(initNetworkDescriptions.size());
-        for (auto& networkDesc : initNetworkDescriptions) {
-            ov::Tensor tensor;
-            tensor = std::move(networkDesc->compiledNetworkTensor);
-
+        initGraphDescriptors.reserve(tensorsInits.size());
+        initNetworkMetadata.reserve(tensorsInits.size());
+        for (const auto& tensor : tensorsInits) {
             GraphDescriptor initGraphDesc;
             NetworkMetadata initNetworkMeta;
             if (_zeGraphExt) {
@@ -203,11 +193,7 @@ std::shared_ptr<IGraph> PluginCompilerAdapter::compileWS(std::shared_ptr<ov::Mod
         std::shared_ptr<ov::Model> targetModel = model;
         size_t i = 0;
 
-        while (auto networkDescription =
-                   std::make_shared<NetworkDescription>(_compiler->compileWsIterative(targetModel, localConfig, i++))) {
-            ov::Tensor tensor;
-            tensor = std::move(networkDescription->compiledNetworkTensor);
-
+        while (auto tensor = _compiler->compileWsIterative(targetModel, localConfig, i++)) {
             GraphDescriptor graphDesc = _zeGraphExt->getGraphDescriptor(tensor.data(), tensor.get_byte_size());
             NetworkMetadata networkMetadata = _zeGraphExt->getNetworkMeta(graphDesc);