[GPU] Enable dynamic quantization gs32 as default for non-systolic (o…

…penvinotoolkit#27119)

### Details:
 - It is applied only to int4 compressed model, non-systolic path
 - Though it is a global configuration, systolic hardware will ignore it

### Tickets:
 - 151708

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

@@ -175,6 +175,7 @@ class debug_configuration {
     } dump_prof_data_iter_params;
 
     static std::ostream* verbose_stream;
+    static const int DYNAMIC_QUANTIZE_GROUP_SIZE_NOT_SET = -2;
 };
 
 }  // namespace cldnn

src/plugins/intel_gpu/src/kernel_selector/kernels/fully_connected/fully_connected_kernel_bf_tiled.cpp

@@ -55,7 +55,7 @@ static size_t get_dynamic_quantize_group_size(const fully_connected_params& para
     auto dynamic_quantization_group_size = params.dynamic_quantization_group_size;
 
     GPU_DEBUG_GET_INSTANCE(debug_config);
-    GPU_DEBUG_IF(debug_config->dynamic_quantize_group_size) {
+    GPU_DEBUG_IF(debug_config->dynamic_quantize_group_size != debug_config->DYNAMIC_QUANTIZE_GROUP_SIZE_NOT_SET) {
         dynamic_quantization_group_size = debug_config->dynamic_quantize_group_size;
 
         // Specify which Fully-connected layer would be dynamic-quantized

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

@@ -872,7 +872,7 @@ void TransformationsPipeline::apply(std::shared_ptr<ov::Model> func) {
         manager.register_pass<ov::pass::Validate>();
 
         auto dynamic_quantization_group_size = config.get_property(ov::hint::dynamic_quantization_group_size);
-        if (device_info.supports_immad) { // XXX: 1048576 is considered per-token
+        if (device_info.supports_immad) {
             pass_config->set_callback<ov::intel_gpu::DynamicQuantizeFullyConnected>([=](const_node_ptr& root) -> bool {
                 if (root->get_input_node_shared_ptr(0)->get_element_type() == ov::element::Type_t::f32) {
                     GPU_DEBUG_TRACE << root->get_friendly_name() << "  Dynamic quantization is turned off because input type is not supported" << std::endl;

src/plugins/intel_gpu/src/runtime/debug_configuration.cpp

@@ -253,7 +253,7 @@ debug_configuration::debug_configuration()
         , disable_runtime_skip_reorder(0)
         , disable_primitive_fusing(0)
         , disable_fake_alignment(0)
-        , dynamic_quantize_group_size(0)
+        , dynamic_quantize_group_size(DYNAMIC_QUANTIZE_GROUP_SIZE_NOT_SET)
         , disable_horizontal_fc_fusion(0) {
 #ifdef GPU_DEBUG_CONFIG
     get_gpu_debug_env_var("Help", help);

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

@@ -46,7 +46,6 @@ void ExecutionConfig::set_default() {
         std::make_tuple(ov::hint::execution_mode, ov::hint::ExecutionMode::PERFORMANCE),
         std::make_tuple(ov::hint::num_requests, 0),
         std::make_tuple(ov::hint::enable_cpu_pinning, false),
-        std::make_tuple(ov::hint::dynamic_quantization_group_size, 0),
 
         std::make_tuple(ov::intel_gpu::hint::host_task_priority, ov::hint::Priority::MEDIUM),
         std::make_tuple(ov::intel_gpu::hint::queue_throttle, ov::intel_gpu::hint::ThrottleLevel::MEDIUM),
@@ -58,7 +57,7 @@ void ExecutionConfig::set_default() {
         std::make_tuple(ov::internal::query_model_ratio, 1.0f),
         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::dynamic_quantization_group_size, 32),
         std::make_tuple(ov::intel_gpu::hint::enable_kernels_reuse, false),
         std::make_tuple(ov::weights_path, ""),
 
@@ -204,7 +203,7 @@ void ExecutionConfig::apply_debug_options(const cldnn::device_info& info) {
         set_property(ov::intel_gpu::use_only_static_kernels_for_dynamic_shape(true));
     }
 
-    GPU_DEBUG_IF(debug_config->dynamic_quantize_group_size) {
+    GPU_DEBUG_IF(debug_config->dynamic_quantize_group_size != debug_config->DYNAMIC_QUANTIZE_GROUP_SIZE_NOT_SET) {
         if (debug_config->dynamic_quantize_group_size == -1)
             set_property(ov::hint::dynamic_quantization_group_size(UINT64_MAX));
         else

src/plugins/intel_gpu/tests/unit/fusions/fully_connected_fusion_test.cpp

@@ -666,6 +666,7 @@ TEST_P(fc_compressed_int8_bias_dynamic_onednn, basic) {
 
     bool is_dynamic = true;
     cfg_not_fused.set_property(ov::intel_gpu::allow_new_shape_infer(is_dynamic));
+    cfg_not_fused.set_property(ov::hint::dynamic_quantization_group_size(0));
     tolerance = 1.0f;
     execute(p, false, is_dynamic);
 }
@@ -705,6 +706,7 @@ TEST_P(fc_compressed_int8_bias_prod_unfused_dynamic_onednn, basic) {
 
     bool is_dynamic = true;
     cfg_not_fused.set_property(ov::intel_gpu::allow_new_shape_infer(is_dynamic));
+    cfg_not_fused.set_property(ov::hint::dynamic_quantization_group_size(0));
     tolerance = 1.0f;
     execute(p, false, is_dynamic);
 }

src/plugins/intel_gpu/tests/unit/test_cases/fully_connected_gpu_test.cpp

@@ -1590,6 +1590,7 @@ class fully_connected_gpu_tests: public ::testing::Test {
             config.set_property(ov::intel_gpu::allow_new_shape_infer(true));
             ov::intel_gpu::ImplementationDesc fc_impl_desc = { format::bfyx, "fully_connected_gpu_bfyx_ref", impl_types::ocl };
             config.set_property(ov::intel_gpu::force_implementations(ov::intel_gpu::ImplForcingMap{ {"fc_prim", fc_impl_desc} }));
+            config.set_property(ov::hint::dynamic_quantization_group_size(0));
 
             network network(engine, topology, config);
             network.set_input_data("input", input_mem);
@@ -1615,6 +1616,7 @@ class fully_connected_gpu_tests: public ::testing::Test {
         auto config = get_test_default_config(engine);
         config.set_property(ov::intel_gpu::allow_new_shape_infer(true));
         config.set_property(ov::intel_gpu::optimize_data(true));
+        config.set_property(ov::hint::dynamic_quantization_group_size(0));
 
         network::ptr network = get_network(engine, topology, config, get_test_stream_ptr(), is_caching_test);
 
@@ -1698,9 +1700,7 @@ class fully_connected_gpu_tests: public ::testing::Test {
             config.set_property(ov::intel_gpu::allow_new_shape_infer(true));
             ov::intel_gpu::ImplementationDesc fc_impl_desc = { format::bfyx, "fully_connected_gpu_bfyx_ref", impl_types::ocl };
             config.set_property(ov::intel_gpu::force_implementations(ov::intel_gpu::ImplForcingMap{ {"fc_prim", fc_impl_desc} }));
-            if (is_dyn_quan) {
-                config.set_property(ov::hint::dynamic_quantization_group_size(0));
-            }
+            config.set_property(ov::hint::dynamic_quantization_group_size(0));
 
             network network(engine, topology, config);
             network.set_input_data("input", input_mem);
@@ -1728,6 +1728,8 @@ class fully_connected_gpu_tests: public ::testing::Test {
         config.set_property(ov::intel_gpu::optimize_data(true));
         if (is_dyn_quan) {
             config.set_property(ov::hint::dynamic_quantization_group_size(32));
+        } else {
+            config.set_property(ov::hint::dynamic_quantization_group_size(0));
         }
 
         network::ptr network = get_network(engine, topology, config, get_test_stream_ptr(), is_caching_test);
@@ -1868,6 +1870,7 @@ class fully_connected_gpu_tests: public ::testing::Test {
             config.set_property(ov::intel_gpu::allow_new_shape_infer(true));
             ov::intel_gpu::ImplementationDesc fc_impl = { in_layout.format, "", impl_types::ocl };
             config.set_property(ov::intel_gpu::force_implementations(ov::intel_gpu::ImplForcingMap{ { "fc_prim1", fc_impl }, { "fc_prim2", fc_impl }  }));
+            config.set_property(ov::hint::dynamic_quantization_group_size(0));
 
             network network(engine, topology, config);
             network.set_input_data("input", input_mem);
@@ -1896,6 +1899,7 @@ class fully_connected_gpu_tests: public ::testing::Test {
         auto config = get_test_default_config(engine);
         config.set_property(ov::intel_gpu::allow_new_shape_infer(true));
         config.set_property(ov::intel_gpu::optimize_data(true));
+        config.set_property(ov::hint::dynamic_quantization_group_size(0));
 
         network::ptr network = get_network(engine, topology, config, get_test_stream_ptr(), is_caching_test);