diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/activation.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/activation.cpp
index d58d19e2b1018a..0cda4e14905d06 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/activation.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/activation.cpp
@@ -73,8 +73,13 @@ struct activation_impl : typed_primitive_impl_ocl<activation> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/arg_max_min.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/arg_max_min.cpp
index d7750e85828e99..a1a869f131b3f4 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/arg_max_min.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/arg_max_min.cpp
@@ -119,8 +119,13 @@ struct arg_max_min_impl : typed_primitive_impl_ocl<arg_max_min> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/border.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/border.cpp
index b8f341f7442ac8..5efc6fdab36b62 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/border.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/border.cpp
@@ -105,8 +105,13 @@ struct border_impl : typed_primitive_impl_ocl<border> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 
     void save(BinaryOutputBuffer& ob) const override {
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/concatenation.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/concatenation.cpp
index 50e48f66df5f05..811a6d968bef31 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/concatenation.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/concatenation.cpp
@@ -82,8 +82,13 @@ struct concatenation_impl : typed_primitive_impl_ocl<concatenation> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/convolution.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/convolution.cpp
index 2c05c3c8ae26d5..86b1071ad7d077 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/convolution.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/convolution.cpp
@@ -261,8 +261,13 @@ struct convolution_impl : typed_primitive_impl_ocl<convolution> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-       auto kernel_params = get_kernel_params(impl_param, true);
-       (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/crop.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/crop.cpp
index c4b11b4ababd3c..6fc4d5dd0dfa9e 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/crop.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/crop.cpp
@@ -47,16 +47,22 @@ struct crop_impl : typed_primitive_impl_ocl<crop> {
         }
         return params;
     }
-        void update_dispatch_data(const kernel_impl_params& impl_param) override {
-            auto kernel_params = get_kernel_params(impl_param, true);
-            auto runtime_offset = convert_data_tensor(impl_param.get_input_layout(), impl_param.input_offsets[0]).GetFirstElementOffset();
-            kernel_selector::ScalarDescriptor s;
-            s.t = kernel_selector::ScalarDescriptor::Types::UINT32;
-            s.v.u32 = static_cast<uint32_t>(runtime_offset);
-            OPENVINO_ASSERT(_kernel_data.kernels[0].params.scalars.size() == 1,
-                    "[GPU] Scalar field for runtime offset is not added for crop shape agnostic impl");
-            _kernel_data.kernels[0].params.scalars[0] = s;
-            (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+
+    void update_dispatch_data(const kernel_impl_params& impl_param) override {
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        auto runtime_offset = convert_data_tensor(impl_param.get_input_layout(), impl_param.input_offsets[0]).GetFirstElementOffset();
+        kernel_selector::ScalarDescriptor s;
+        s.t = kernel_selector::ScalarDescriptor::Types::UINT32;
+        s.v.u32 = static_cast<uint32_t>(runtime_offset);
+        OPENVINO_ASSERT(_kernel_data.kernels[0].params.scalars.size() == 1,
+                "[GPU] Scalar field for runtime offset is not added for crop shape agnostic impl");
+        _kernel_data.kernels[0].params.scalars[0] = s;
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/cum_sum.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/cum_sum.cpp
index 414d70e3b03c2f..c806a785c530f2 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/cum_sum.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/cum_sum.cpp
@@ -78,8 +78,13 @@ struct cum_sum_impl : typed_primitive_impl_ocl<cum_sum> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/fully_connected.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/fully_connected.cpp
index 96bafc7e431f3b..d3acb9dd6a9b55 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/fully_connected.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/fully_connected.cpp
@@ -77,7 +77,7 @@ struct fully_connected_impl : typed_primitive_impl_ocl<fully_connected> {
     }
 
 public:
-    static kernel_params_t get_kernel_params(const kernel_impl_params& impl_param, bool is_shape_agnostic = false) {
+    static kernel_impl_params update_impl_params(const kernel_impl_params& impl_param) {
         const auto& primitive = impl_param.typed_desc<fully_connected>();
 
         auto get_fc_input_layouts = [primitive](const std::vector<layout>& input_layouts, bool allow_new_shape_infer) {
@@ -151,12 +151,19 @@ struct fully_connected_impl : typed_primitive_impl_ocl<fully_connected> {
         auto updated_impl_param = impl_param;
 
         const auto input_layouts = get_fc_input_layouts(impl_param.input_layouts, allow_new_shape_infer);
-        updated_impl_param.input_layouts[0] = input_layouts[0];
-        updated_impl_param.input_layouts[1] = input_layouts[1];
+        for (size_t i = 0; i < input_layouts.size(); ++i) {
+            updated_impl_param.input_layouts[i] = input_layouts[i];
+        }
         updated_impl_param.weights_layout = input_layouts[1];
 
         updated_impl_param.output_layouts[0] = get_fc_output_layout(input_layouts, impl_param.get_output_layout());
 
+        return updated_impl_param;
+    }
+
+    static kernel_params_t get_kernel_params(const kernel_impl_params& impl_param, bool is_shape_agnostic = false) {
+        const auto& primitive = impl_param.typed_desc<fully_connected>();
+        auto updated_impl_param = update_impl_params(impl_param);
         auto params = get_weights_bias_default_params<kernel_selector::fully_connected_params>(updated_impl_param, false, is_shape_agnostic);
         params.allowInputReordering = true;
 
@@ -164,10 +171,10 @@ struct fully_connected_impl : typed_primitive_impl_ocl<fully_connected> {
         bool with_zp = !primitive->decompression_zero_point.empty();
         if (commpressed) {
             params.compressed = true;
-            params.decompression_scale = convert_data_tensor(input_layouts[2]);
+            params.decompression_scale = convert_data_tensor(updated_impl_param.input_layouts[2]);
             if (with_zp) {
                 params.has_decompression_zp = true;
-                params.decompression_zero_point = convert_data_tensor(input_layouts[3]);
+                params.decompression_zero_point = convert_data_tensor(updated_impl_param.input_layouts[3]);
             } else if (primitive->decompression_zero_point_scalar.has_value()) {
                 params.has_decompression_zp = true;
                 params.scalar_zp = true;
@@ -194,8 +201,20 @@ struct fully_connected_impl : typed_primitive_impl_ocl<fully_connected> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        auto& params = static_cast<kernel_params_t&>(*_kernel_data.params);
+        auto updated_impl_param = update_impl_params(impl_param);
+        update_shapes(params, updated_impl_param);
+
+        if (impl_param.typed_desc<fully_connected>()->input_size != 3) {
+            params.outputs = { params.outputs[0].FlattenFeatureAndSpatials() };
+        }
+
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/gather.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/gather.cpp
index af69439d7aa51a..0d20294c6bc33b 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/gather.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/gather.cpp
@@ -167,8 +167,13 @@ struct gather_impl : typed_primitive_impl_ocl<gather> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/gather_elements.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/gather_elements.cpp
index 317db10b1463b3..6a683090a7f4ac 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/gather_elements.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/gather_elements.cpp
@@ -75,8 +75,13 @@ struct gather_elements_impl : typed_primitive_impl_ocl<gather_elements> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-       auto kernel_params = get_kernel_params(impl_param, true);
-       (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/gather_nd.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/gather_nd.cpp
index ac7c8da1ac15eb..8ea57b56614cc9 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/gather_nd.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/gather_nd.cpp
@@ -45,8 +45,13 @@ struct gather_nd_impl : typed_primitive_impl_ocl<gather_nd> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/gemm.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/gemm.cpp
index f7220dad387348..979d1efb8cd76c 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/gemm.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/gemm.cpp
@@ -303,12 +303,19 @@ struct gemm_impl : multi_stage_primitive<gemm> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true, false);
-        (_kernels_data[default_gemm].update_dispatch_data_func)(kernel_params, _kernels_data[default_gemm]);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernels_data[default_gemm].params == nullptr) {
+            _kernels_data[default_gemm].params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true, false));
+        }
+        update_shapes(*_kernels_data[default_gemm].params, impl_param);
+        (_kernels_data[default_gemm].update_dispatch_data_func)(*_kernels_data[default_gemm].params, _kernels_data[default_gemm]);
 
         if (_kernels_data.size() == 2) {
-            auto kernel_params = get_kernel_params(impl_param, true, true);
-            (_kernels_data[indirect_gemm].update_dispatch_data_func)(kernel_params, _kernels_data[indirect_gemm]);
+            if (_kernels_data[indirect_gemm].params == nullptr) {
+                _kernels_data[indirect_gemm].params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true, true));
+            }
+            update_shapes(*_kernels_data[indirect_gemm].params, impl_param);
+            (_kernels_data[indirect_gemm].update_dispatch_data_func)(*_kernels_data[indirect_gemm].params, _kernels_data[indirect_gemm]);
         }
     }
 };
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/group_normalization.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/group_normalization.cpp
index 827f05c2d2e20f..d79e47e8a114e0 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/group_normalization.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/group_normalization.cpp
@@ -32,8 +32,13 @@ struct group_normalization_impl : typed_primitive_impl_ocl<group_normalization>
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/multinomial.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/multinomial.cpp
index bcc2fa15be20b3..45607326ff2925 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/multinomial.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/multinomial.cpp
@@ -32,8 +32,13 @@ struct multinomial_impl : typed_primitive_impl_ocl<multinomial> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/mvn.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/mvn.cpp
index c370bd8dff4c40..a3de617405fbad 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/mvn.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/mvn.cpp
@@ -88,8 +88,13 @@ struct mvn_impl : typed_primitive_impl_ocl<mvn> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/non_max_suppression.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/non_max_suppression.cpp
index 81194cb8ba9c5a..65bfa94173bf11 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/non_max_suppression.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/non_max_suppression.cpp
@@ -60,8 +60,9 @@ struct non_max_suppression_impl : typed_primitive_impl_ocl<non_max_suppression>
     }
 
 public:
-    static std::unique_ptr<primitive_impl> create(const non_max_suppression_node& arg, const kernel_impl_params& impl_param) {
+static kernel_params_t get_kernel_params(const kernel_impl_params& impl_param, bool is_shape_agnostic = false) {
         const auto& primitive = impl_param.typed_desc<non_max_suppression>();
+        const auto& arg = impl_param.prog->get_node(impl_param.desc->id).as<non_max_suppression>();
         auto params = get_default_params<kernel_selector::non_max_suppression_params>(impl_param);
 
         const auto input_scores_idx = 1;
@@ -154,11 +155,7 @@ struct non_max_suppression_impl : typed_primitive_impl_ocl<non_max_suppression>
             params.reuse_internal_buffer = true;
         }
 
-        params.set_dynamic_shape_offsets();
-        auto& kernel_selector = kernel_selector::non_max_suppression_kernel_selector::Instance();
-        auto best_kernel = kernel_selector.get_best_kernel(params);
-
-        return make_unique<non_max_suppression_impl>(best_kernel);
+        return params;
     }
 
 private:
@@ -200,7 +197,7 @@ namespace detail {
 
 attach_non_max_suppression_impl::attach_non_max_suppression_impl() {
     implementation_map<non_max_suppression>::add(impl_types::ocl,
-                                                 non_max_suppression_impl::create,
+                                                 typed_primitive_impl_ocl<non_max_suppression>::create<non_max_suppression_impl>,
                                                  {
                                                      std::make_tuple(data_types::i32, format::bfyx),
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/non_zero.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/non_zero.cpp
index c8f0a5cbae10c0..d8f0e45c25146f 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/non_zero.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/non_zero.cpp
@@ -48,8 +48,13 @@ struct count_nonzero_impl : typed_primitive_impl_ocl<count_nonzero> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
@@ -82,8 +87,13 @@ struct gather_nonzero_impl : typed_primitive_impl_ocl<gather_nonzero> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/permute.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/permute.cpp
index 06ea0a215d12cf..e253fef751b091 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/permute.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/permute.cpp
@@ -71,8 +71,13 @@ struct permute_impl : typed_primitive_impl_ocl<permute> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/quantize.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/quantize.cpp
index f1517b13824c42..73a668966ceeb5 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/quantize.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/quantize.cpp
@@ -54,8 +54,9 @@ struct quantize_impl : typed_primitive_impl_ocl<quantize> {
     }
 
 public:
-    static std::unique_ptr<primitive_impl> create(const quantize_node& arg, const kernel_impl_params& impl_param) {
+    static kernel_params_t get_kernel_params(const kernel_impl_params& impl_param, bool is_shape_agnostic = false) {
         auto quantize_params = get_default_params<kernel_selector::quantize_params>(impl_param);
+        const auto& arg = impl_param.prog->get_node(impl_param.desc->id).as<quantize>();
 
         quantize_params.levels = arg.get_levels();
         quantize_params.scale_shift_opt = arg.get_scale_shift_opt();
@@ -86,17 +87,17 @@ struct quantize_impl : typed_primitive_impl_ocl<quantize> {
             quantize_params.inputs.push_back(convert_data_tensor(impl_param.input_layouts[i]));
         }
 
-        quantize_params.is_shape_agnostic = impl_param.is_dynamic();
-        quantize_params.set_dynamic_shape_offsets();
-        auto& kernel_selector = kernel_selector::quantize_kernel_selector::Instance();
-        auto best_kernel = kernel_selector.get_best_kernel(quantize_params);
-
-        return make_unique<quantize_impl>(best_kernel);
+        return quantize_params;
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto quantize_params = get_default_params<kernel_selector::quantize_params>(impl_param);
-        (_kernel_data.update_dispatch_data_func)(quantize_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
@@ -147,9 +148,16 @@ attach_quantize_impl::attach_quantize_impl() {
     keys.emplace(data_types::f16, format::yxfb);
     keys.emplace(data_types::f32, format::yxfb);
 
-    implementation_map<quantize>::add(impl_types::ocl, shape_types::static_shape, quantize_impl::create, keys);
+    implementation_map<quantize>::add(impl_types::ocl,
+                                      shape_types::static_shape,
+                                      typed_primitive_impl_ocl<quantize>::create<quantize_impl>,
+                                      keys);
 
-    implementation_map<quantize>::add(impl_types::ocl, shape_types::dynamic_shape, quantize_impl::create, types, dyn_formats);
+    implementation_map<quantize>::add(impl_types::ocl,
+                                      shape_types::dynamic_shape,
+                                      typed_primitive_impl_ocl<quantize>::create<quantize_impl>,
+                                      types,
+                                      dyn_formats);
 }
 
 }  // namespace detail
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/reduce.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/reduce.cpp
index c004b63c4f23fd..d937d7fc8a190c 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/reduce.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/reduce.cpp
@@ -90,8 +90,13 @@ struct reduce_impl : typed_primitive_impl_ocl<reduce> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/reorder.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/reorder.cpp
index 02d2e9e19827ec..aa030af71dfc3d 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/reorder.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/reorder.cpp
@@ -119,8 +119,13 @@ struct reorder_impl : typed_primitive_impl_ocl<reorder> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 
     static std::unique_ptr<primitive_impl> create(const reorder_node& arg, const kernel_impl_params& impl_param) {
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/rms.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/rms.cpp
index 946334f7c59d90..34e379564ee10e 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/rms.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/rms.cpp
@@ -43,8 +43,13 @@ struct rms_impl : typed_primitive_impl_ocl<rms> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 
     static kernel_impl_params static_canonicalize_shapes(const kernel_impl_params& impl_params) {
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/rope.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/rope.cpp
index 4706227e35cff2..f65768b8e6eb20 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/rope.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/rope.cpp
@@ -91,8 +91,13 @@ struct rope_impl : typed_primitive_impl_ocl<rope> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/scaled_dot_product_attention.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/scaled_dot_product_attention.cpp
index 4bc6e16ce55fa9..fd2270fa366ec6 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/scaled_dot_product_attention.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/scaled_dot_product_attention.cpp
@@ -275,12 +275,19 @@ struct scaled_dot_product_attention_impl : multi_stage_primitive<scaled_dot_prod
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-       auto kernel_params = get_kernel_params(impl_param, true);
-       (_kernels_data[default_sdpa].update_dispatch_data_func)(kernel_params, _kernels_data[default_sdpa]);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernels_data[default_sdpa].params == nullptr) {
+            _kernels_data[default_sdpa].params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+        update_shapes(*_kernels_data[default_sdpa].params, impl_param);
+        (_kernels_data[default_sdpa].update_dispatch_data_func)(*_kernels_data[default_sdpa].params, _kernels_data[default_sdpa]);
 
         if (_kernels_data.size() == 2) {
-            auto kernel_params = get_kernel_params(impl_param, true);
-            (_kernels_data[indirect_sdpa].update_dispatch_data_func)(kernel_params, _kernels_data[indirect_sdpa]);
+            if (_kernels_data[indirect_sdpa].params == nullptr) {
+                _kernels_data[indirect_sdpa].params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+            }
+            update_shapes(*_kernels_data[indirect_sdpa].params, impl_param);
+            (_kernels_data[indirect_sdpa].update_dispatch_data_func)(*_kernels_data[indirect_sdpa].params, _kernels_data[indirect_sdpa]);
         }
     }
 };
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/scatter_nd_update.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/scatter_nd_update.cpp
index 8709b8be75c0ea..e5f1ebd04b1d4a 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/scatter_nd_update.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/scatter_nd_update.cpp
@@ -45,8 +45,13 @@ struct scatter_nd_update_impl : typed_primitive_impl_ocl<scatter_nd_update> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-       auto kernel_params = get_kernel_params(impl_param, true);
-       (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/scatter_update.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/scatter_update.cpp
index 3e59ecbf63691e..af1029aacb2036 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/scatter_update.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/scatter_update.cpp
@@ -70,8 +70,13 @@ struct scatter_update_impl : typed_primitive_impl_ocl<scatter_update> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-       auto kernel_params = get_kernel_params(impl_param, true);
-       (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/select.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/select.cpp
index 90a369006715d7..0f69ab377ed3fe 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/select.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/select.cpp
@@ -61,8 +61,13 @@ struct select_impl : typed_primitive_impl_ocl<select> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/shape_of.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/shape_of.cpp
index ee17337d39044f..70e20b1a01ed53 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/shape_of.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/shape_of.cpp
@@ -59,8 +59,13 @@ struct shape_of_impl : typed_primitive_impl_ocl<shape_of> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/slice.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/slice.cpp
index 175e26b1a3c5d8..8a7225cb3c64ce 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/slice.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/slice.cpp
@@ -88,9 +88,10 @@ struct slice_impl : typed_primitive_impl_ocl<slice> {
         return args;
     }
 
-    static std::unique_ptr<primitive_impl> create(const slice_node& arg, const kernel_impl_params& impl_param) {
-        auto params = get_default_params<kernel_selector::slice_params>(impl_param, impl_param.is_dynamic());
+    static kernel_params_t get_kernel_params(const kernel_impl_params& impl_param, bool is_shape_agnostic = false) {
+        auto params = get_default_params<kernel_selector::slice_params>(impl_param, is_shape_agnostic);
         const auto input_rank = params.inputs[0].Dimentions();
+        const auto& arg = impl_param.prog->get_node(impl_param.desc->id);
 
         if (!PrepareInput(arg,
                           SliceKernelRefNeededInputs::kStart,
@@ -135,15 +136,18 @@ struct slice_impl : typed_primitive_impl_ocl<slice> {
         }
 
         params.set_dynamic_shape_offsets();
-        auto& kernel_selector = kernel_selector::slice_kernel_selector::Instance();
-        auto best_kernel = kernel_selector.get_best_kernel(params);
 
-        return make_unique<slice_impl>(best_kernel);
+        return params;
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_default_params<kernel_selector::slice_params>(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 
 private:
@@ -177,18 +181,28 @@ struct slice_impl : typed_primitive_impl_ocl<slice> {
 namespace detail {
 
 attach_slice_impl::attach_slice_impl() {
-    auto types = {data_types::f32, data_types::f16, data_types::i8, data_types::u8, data_types::i32, data_types::i64};
+    auto types = {
+        data_types::f32,
+        data_types::f16,
+        data_types::i8,
+        data_types::u8,
+        data_types::i32,
+        data_types::i64
+    };
 
     auto formats = {
         format::bfyx,
         format::bfzyx,
     };
 
-    implementation_map<slice>::add(impl_types::ocl, shape_types::any, slice_impl::create, types, formats);
+    implementation_map<slice>::add(impl_types::ocl,
+                                   shape_types::any,
+                                   typed_primitive_impl_ocl<slice>::create<slice_impl>,
+                                   types,
+                                   formats);
 }
 
 }  // namespace detail
-
 }  // namespace ocl
 }  // namespace cldnn
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/space_to_batch.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/space_to_batch.cpp
index c9800619849b90..75568cc8984bd3 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/space_to_batch.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/space_to_batch.cpp
@@ -59,8 +59,13 @@ struct space_to_batch_impl : typed_primitive_impl_ocl<space_to_batch> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/swiglu.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/swiglu.cpp
index be77e13a0f21bb..8f96f7afa31556 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/swiglu.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/swiglu.cpp
@@ -46,8 +46,13 @@ struct swiglu_impl : typed_primitive_impl_ocl<swiglu> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/tile.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/tile.cpp
index 0dd1636213e7e8..6d2268de0730e7 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/tile.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/tile.cpp
@@ -52,8 +52,13 @@ struct tile_impl : typed_primitive_impl_ocl<tile> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/unique.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/unique.cpp
index ce1e32944a760c..d90580de3c35fd 100644
--- a/src/plugins/intel_gpu/src/graph/impls/ocl/unique.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/ocl/unique.cpp
@@ -42,8 +42,13 @@ struct unique_count_impl : typed_primitive_impl_ocl<unique_count> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };
 
@@ -128,8 +133,13 @@ struct unique_gather_impl : typed_primitive_impl_ocl<unique_gather> {
     }
 
     void update_dispatch_data(const kernel_impl_params& impl_param) override {
-        auto kernel_params = get_kernel_params(impl_param, true);
-        (_kernel_data.update_dispatch_data_func)(kernel_params, _kernel_data);
+        // If model loaded from cache, params are not initialized, so we create a new object and reuse it in the future
+        if (_kernel_data.params == nullptr) {
+            _kernel_data.params = std::make_shared<kernel_params_t>(get_kernel_params(impl_param, true));
+        }
+
+        update_shapes(*_kernel_data.params, impl_param);
+        (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 };