Fixes dynamic shapes (#27776)

### Details:

* Aligns fixes in dynamic shape serializing with
intel-innersource/applications.ai.vpu-accelerators.vpux-plugin#14490
* Fix parsing of shapes for SIT

### Tickets:
 - E147314
 - E147315

---------

Signed-off-by: Skrebkov, Artemy <[email protected]>
Co-authored-by: Pawel Raasz <[email protected]>

samples/cpp/benchmark_app/main.cpp

@@ -523,9 +523,7 @@ int main(int argc, char* argv[]) {
             }
         }
         auto result = std::find_if(config.begin(), config.end(), [&](const std::pair<std::string, ov::AnyMap>& item) {
-            if (device_name.find(item.first) == 0)
-                return true;
-            return false;
+            return device_name.find(item.first) == 0;
         });
         ov::AnyMap device_config = {};
         if (result != config.end())
@@ -548,6 +546,11 @@ int main(int argc, char* argv[]) {
         }
 
         bool isDynamicNetwork = false;
+        auto areNetworkInputsDynamic = [](const benchmark_app::InputsInfo& input_info) {
+            return std::any_of(input_info.begin(), input_info.end(), [](const auto& info) {
+                return info.second.partialShape.is_dynamic();
+            });
+        };
 
         if (FLAGS_load_from_file && !isNetworkCompiled) {
             if (!FLAGS_mean_values.empty() || !FLAGS_scale_values.empty()) {
@@ -722,12 +725,7 @@ int main(int argc, char* argv[]) {
             model = preproc.build();
 
             // Check if network has dynamic shapes
-            auto input_info = app_inputs_info[0];
-            isDynamicNetwork = std::any_of(input_info.begin(),
-                                           input_info.end(),
-                                           [](const std::pair<std::string, benchmark_app::InputInfo>& i) {
-                                               return i.second.partialShape.is_dynamic();
-                                           });
+            isDynamicNetwork = areNetworkInputsDynamic(app_inputs_info.at(0));
 
             topology_name = model->get_friendly_name();
 
@@ -789,6 +787,7 @@ int main(int argc, char* argv[]) {
                                               FLAGS_scale_values,
                                               FLAGS_mean_values,
                                               compiledModel.inputs());
+            isDynamicNetwork = areNetworkInputsDynamic(app_inputs_info.at(0));
 
             batchSize = get_batch_size(app_inputs_info.at(0));
             warn_if_no_batch(app_inputs_info.at(0));

src/plugins/intel_npu/src/backend/src/zero_infer_request.cpp

@@ -54,9 +54,8 @@ void check_level_zero_attributes_match(const IODescriptor& ioDescriptor, const A
                         '\n' + "Given: " + std::to_string(ovDimensions.size()));
 
     for (size_t index = 0; index < ovDimensions.size(); ++index) {
-        OPENVINO_ASSERT(
-            ioDescriptor.shapeFromCompiler.is_dynamic() || ovDimensions[index] == zeDescriptor.info.dims[index],
-            "Shape mismatch for input/output named " + ioDescriptor.nameFromCompiler);
+        OPENVINO_ASSERT(ovDimensions[index] == zeDescriptor.info.dims[index],
+                        "Shape mismatch for input/output named " + ioDescriptor.nameFromCompiler);
     }
     for (size_t index = ovDimensions.size(); index < ZE_MAX_GRAPH_ARGUMENT_DIMENSIONS_SIZE; ++index) {
         OPENVINO_ASSERT(zeDescriptor.info.dims[index] == 0 || zeDescriptor.info.dims[index] == 1,

src/plugins/intel_npu/src/compiler_adapter/src/ze_graph_ext_wrappers.cpp

@@ -12,7 +12,9 @@
 #include "intel_npu/utils/zero/zero_api.hpp"
 #include "intel_npu/utils/zero/zero_result.hpp"
 #include "intel_npu/utils/zero/zero_wrappers.hpp"
+#include "openvino/core/dimension.hpp"
 #include "openvino/core/model.hpp"
+#include "openvino/core/partial_shape.hpp"
 
 #define NotSupportQuery(T) (T <= ZE_GRAPH_EXT_VERSION_1_2)
 
@@ -400,7 +402,8 @@ ze_graph_handle_t ZeGraphExtWrappers::getGraphHandle(const std::vector<uint8_t>&
 static IODescriptor getIODescriptor(const ze_graph_argument_properties_3_t& arg,
                                     const std::optional<ze_graph_argument_metadata_t>& metadata) {
     ov::element::Type_t precision = toOVElementType(arg.devicePrecision);
-    ov::Shape shapeFromCompiler, shapeFromIRModel;
+    ov::Shape shapeFromCompiler;
+    ov::PartialShape shapeFromIRModel;
     std::unordered_set<std::string> outputTensorNames;
 
     for (uint32_t id = 0; id < arg.associated_tensor_names_count; id++) {
@@ -410,8 +413,17 @@ static IODescriptor getIODescriptor(const ze_graph_argument_properties_3_t& arg,
         shapeFromCompiler.push_back(arg.dims[id]);
     }
     if (metadata.has_value()) {
+        const auto dynamicDim = std::numeric_limits<uint64_t>::max();
+        shapeFromIRModel.reserve(metadata->shape_size);
         for (uint32_t id = 0; id < metadata->shape_size; id++) {
-            shapeFromIRModel.push_back(metadata->shape[id]);
+            if (metadata->shape[id] != dynamicDim) {
+                shapeFromIRModel.push_back(metadata->shape[id]);
+            } else {
+                // lower bound is ignored, so we set it to 1 just to satisfy the Dimension constructor,
+                // upper bound is set to the value from shapeFromCompiler as it is filled with upper bounds
+                // in case of dynamic dimensions
+                shapeFromIRModel.push_back(ov::Dimension(1, shapeFromCompiler[id]));
+            }
         }
     }
 
@@ -433,7 +445,7 @@ static IODescriptor getIODescriptor(const ze_graph_argument_properties_3_t& arg,
 
     return {std::move(nameFromCompiler),
             precision,
-            std::move(shapeFromCompiler),
+            shapeFromCompiler,
             isStateInput,
             isStateOutput,
             isShapeTensor,

src/plugins/intel_npu/tools/single-image-test/main.cpp

@@ -1569,8 +1569,8 @@ std::pair<TensorMap, ProfVec> runInfer(ov::InferRequest& inferRequest, ov::Compi
 
     TensorMap out;
     for (const auto& outputInfo : compiledModel.outputs()) {
-        const std::string layer_name = outputInfo.get_any_name();
-        out.insert({layer_name, inferRequest.get_tensor(layer_name)});
+        const std::string layerName = outputInfo.get_any_name();
+        out.insert({layerName, inferRequest.get_tensor(layerName)});
     }
 
     ProfVec profData{};
@@ -1807,11 +1807,17 @@ bool testMeanIoU(const TensorMap& outputs, const TensorMap& references, const La
 }
 
 static ov::Shape parseDataShape(const std::string& dataShapeStr) {
-    std::vector<size_t> dataShape;
-    std::istringstream ss(dataShapeStr);
-    std::string token;
-    while (std::getline(ss, token, ',')) {
-        dataShape.push_back(std::stoul(token));
+    std::vector<uint64_t> dataShape;
+    std::stringstream ss(dataShapeStr);
+
+    char ch;  // To discard non-numeric characters
+    int64_t dim;
+    while (ss >> ch) {
+        if (std::isdigit(ch)) {
+            ss.putback(ch);
+            ss >> dim;
+            dataShape.push_back(dim);
+        }
     }
     return ov::Shape(dataShape);
 }
@@ -1906,11 +1912,11 @@ static int runSingleImageTest() {
             auto model = core.read_model(FLAGS_network);
             nameIOTensors(model);
 
-            auto inputs_info = std::const_pointer_cast<ov::Model>(model)->inputs();
-            InputsInfo info_map;
+            auto inputsInfo = std::const_pointer_cast<ov::Model>(model)->inputs();
+            InputsInfo infoMap;
 
             std::cout << "Performing reshape" << std::endl;
-            reshape(std::move(inputs_info), info_map, model, FLAGS_shape,
+            reshape(std::move(inputsInfo), infoMap, model, FLAGS_shape,
                     FLAGS_override_model_batch_size, FLAGS_device);
 
             ov::preprocess::PrePostProcessor ppp(model);