[GPU] Changed default value for kv cache precision (#28502)

### Details:
- Set default precision of kv cache as fp16 to match inference precision
- Update logic which handles multiple apply_user_properties() call for
same object to skip everything if config is already finalized.
- For platforms with xmx support set `kv_cache_precision` as
`inference_precision` to provide correct property query result

---------

Signed-off-by: Vladimir Paramuzov <[email protected]>

src/plugins/intel_gpu/include/intel_gpu/runtime/execution_config.hpp

@@ -150,7 +150,6 @@ class ExecutionConfig {
     void apply_performance_hints(const cldnn::device_info& info);
     void apply_priority_hints(const cldnn::device_info& info);
     void apply_debug_options(const cldnn::device_info& info);
-    void update_specific_default_properties(const cldnn::device_info& info);
 
     template <typename T, PropertyMutability mutability>
     void apply_rt_info_property(const ov::Property<T, mutability>& property, const ov::RTMap& rt_info) {
@@ -169,7 +168,7 @@ class ExecutionConfig {
     std::map<std::string, PropertyVisibility> supported_properties;
     std::map<std::string, BaseValidator::Ptr> property_validators;
 
-    bool specific_default_properties_is_set = false;
+    bool finalized = false;
 };
 
 }  // namespace intel_gpu

src/plugins/intel_gpu/src/graph/program.cpp

@@ -530,6 +530,7 @@ void program::init_graph() {
             node->get_output_layouts();
         if (node->is_type<lstm_seq>()) {
             _config.set_property(ov::intel_gpu::use_onednn(true));
+            _config.set_property(ov::intel_gpu::queue_type(QueueTypes::in_order));
         }
     }
     // Perform initial shape_of subgraphs markup

src/plugins/intel_gpu/src/plugin/plugin.cpp

@@ -240,10 +240,12 @@ std::shared_ptr<ov::ICompiledModel> Plugin::compile_model(const std::shared_ptr<
     auto context_impl = get_context_impl(context);
     auto device_id = ov::DeviceIDParser{context_impl->get_device_name()}.get_device_id();
 
-    OPENVINO_ASSERT(m_configs_map.find(device_id) != m_configs_map.end(), "[GPU] LoadExeNetworkImpl: Couldn't find config for GPU with id ", device_id);
+    OPENVINO_ASSERT(m_configs_map.find(device_id) != m_configs_map.end(), "[GPU] compile_model: Couldn't find config for GPU with id ", device_id);
 
     ExecutionConfig config = m_configs_map.at(device_id);
     config.set_user_property(orig_config);
+    if (model->has_rt_info("runtime_options"))
+        config.apply_rt_info(context_impl->get_engine().get_device_info(), model->get_rt_info<ov::AnyMap>("runtime_options"), is_llm(model));
     config.apply_user_properties(context_impl->get_engine().get_device_info());
 
     set_cache_info(model, config);

src/plugins/intel_gpu/src/runtime/execution_config.cpp

@@ -5,6 +5,7 @@
 #include "intel_gpu/runtime/execution_config.hpp"
 #include "intel_gpu/runtime/debug_configuration.hpp"
 #include "openvino/runtime/internal_properties.hpp"
+#include "openvino/runtime/properties.hpp"
 
 #include <thread>
 
@@ -59,7 +60,7 @@ void ExecutionConfig::set_default() {
         std::make_tuple(ov::cache_mode, ov::CacheMode::OPTIMIZE_SPEED),
         std::make_tuple(ov::cache_encryption_callbacks, EncryptionCallbacks{}),
         std::make_tuple(ov::hint::dynamic_quantization_group_size, 0),
-        std::make_tuple(ov::hint::kv_cache_precision, ov::element::undefined),
+        std::make_tuple(ov::hint::kv_cache_precision, ov::element::f16),
         std::make_tuple(ov::intel_gpu::hint::enable_kernels_reuse, false),
         std::make_tuple(ov::weights_path, ""),
         std::make_tuple(ov::hint::activations_scale_factor, -1.f),
@@ -230,26 +231,9 @@ void ExecutionConfig::apply_hints(const cldnn::device_info& info) {
     apply_debug_options(info);
 }
 
-void ExecutionConfig::update_specific_default_properties(const cldnn::device_info& info) {
-    // These default properties should be set once.
-    if (specific_default_properties_is_set)
-        return;
-    specific_default_properties_is_set = true;
-
-    // Enable KV-cache compression by default for non-systolic platforms MFDNN-11755
-    if (get_property(ov::hint::kv_cache_precision) == ov::element::undefined && !info.supports_immad) {
-        set_property(ov::hint::kv_cache_precision(ov::element::i8));
-    }
-
-    // Enable dynamic quantization by default for non-systolic platforms
-    if (get_property(ov::hint::dynamic_quantization_group_size) == 0 && !info.supports_immad) {
-        set_property(ov::hint::dynamic_quantization_group_size(32));
-    }
-}
-
 void ExecutionConfig::apply_user_properties(const cldnn::device_info& info) {
-    // Update specific default properties, call once before internal_properties updated.
-    update_specific_default_properties(info);
+    if (finalized)
+        return;
 
     // Copy internal properties before applying hints to ensure that
     // a property set by hint won't be overriden by a value in user config.
@@ -280,6 +264,23 @@ void ExecutionConfig::apply_user_properties(const cldnn::device_info& info) {
         }
     }
 
+    if (!is_set_by_user(ov::hint::kv_cache_precision) || get_property(ov::hint::kv_cache_precision) == ov::element::undefined) {
+        if (info.supports_immad) {  // MFDNN-11755
+            set_property(ov::hint::kv_cache_precision(get_property(ov::hint::inference_precision)));
+        } else {
+            // Enable KV-cache compression by default for non-systolic platforms only
+            set_property(ov::hint::kv_cache_precision(ov::element::i8));
+        }
+    }
+
+    // Enable dynamic quantization by default for non-systolic platforms
+    if (!is_set_by_user(ov::hint::dynamic_quantization_group_size) &&
+         get_property(ov::hint::dynamic_quantization_group_size) == 0 && !info.supports_immad) {
+        set_property(ov::hint::dynamic_quantization_group_size(32));
+    }
+
+    finalized = true;
+
     user_properties.clear();
 }
 

src/plugins/intel_gpu/tests/functional/subgraph_tests/dynamic/kv_cache.cpp

@@ -539,7 +539,7 @@ class KVCacheIssueTests: public ::testing::Test {
         auto core = ov::test::utils::PluginCache::get().core();
 
         ov::AnyMap properties = {
-            ov::hint::kv_cache_precision(ov::element::undefined)
+            ov::hint::kv_cache_precision(ov::element::f16)
         };
 
         const size_t n_batch = 1;

src/plugins/intel_gpu/tests/functional/subgraph_tests/dynamic/kv_cache_sdpa.cpp

@@ -50,7 +50,7 @@ class SDPAWithKVCacheTest : public ::testing::Test, public ::testing::WithParamI
         if (p.compressed) {
             properties.emplace(ov::hint::kv_cache_precision(ov::element::i8));
         } else {
-            properties.emplace(ov::hint::kv_cache_precision(ov::element::undefined));
+            properties.emplace(ov::hint::kv_cache_precision(ov::element::f16));
         }
 
         const size_t n_heads = 16;