[GPU] Fix incorrect broadcast axis for static SDPA case

Author: sshlyapn

src/plugins/intel_gpu/src/graph/impls/ocl/scaled_dot_product_attention.cpp

@@ -249,12 +249,12 @@ struct scaled_dot_product_attention_impl : multi_stage_primitive<scaled_dot_prod
         const auto value_shape = transpose_pshape(impl_param.get_input_layout(2).get_partial_shape(), desc->input_v_transpose_order);
 
         OPENVINO_ASSERT(key_shape == value_shape, "[GPU] The shapes of key and value inputs are expected to be equal");
-        for (size_t i = 0; i < query_shape.size(); ++i) {
-            if (query_shape[i].is_static() && key_shape[i].is_static() && value_shape[i].is_static()) {
-                if (query_shape[i].get_length() > key_shape[i].get_length()) {
-                    config.broadcast_axis = desc->input_k_transpose_order[i];
-                    config.group_size = query_shape[i].get_length() / key_shape[i].get_length();
-                }
+
+        const auto num_heads_dim = 1;
+        if (query_shape[num_heads_dim].is_static() && key_shape[num_heads_dim].is_static() && value_shape[num_heads_dim].is_static()) {
+            if (query_shape[num_heads_dim].get_length() > key_shape[num_heads_dim].get_length()) {
+                config.broadcast_axis = desc->input_k_transpose_order[num_heads_dim];
+                config.group_size = query_shape[num_heads_dim].get_length() / key_shape[num_heads_dim].get_length();
             }
         }
 

src/plugins/intel_gpu/tests/functional/single_layer_tests/dynamic/scaled_dot_product_attention.cpp

@@ -329,18 +329,63 @@ const std::vector<std::vector<InputShape>> shapes{
 };
 
 const std::vector<std::vector<int64_t>> disable_transpose{};
-const std::vector<std::vector<int64_t>> enable_transpose{{0, 1, 2, 3}, {0, 1, 2, 3}, {0, 2, 1, 3}};
+const std::vector<std::vector<int64_t>> transpose_value{{0, 1, 2, 3}, {0, 1, 2, 3}, {0, 2, 1, 3}};
+const std::vector<std::vector<int64_t>> transpose_all{{0, 2, 1, 3}, {0, 2, 1, 3}, {0, 2, 1, 3}};
 
 const auto params = testing::Combine(testing::Values(ov::element::f16 /*, ov::element::f32 */),
-                                                 testing::ValuesIn(shapes),
-                                                 testing::Values(true, false),
-                                                 testing::Values(true, false),
-                                                 testing::Values(true, false),
-                                                 testing::ValuesIn({disable_transpose, enable_transpose}));
+                                     testing::ValuesIn(shapes),
+                                     testing::Values(true, false),
+                                     testing::Values(true, false),
+                                     testing::Values(true, false),
+                                     testing::ValuesIn({disable_transpose, transpose_value}));
 
 INSTANTIATE_TEST_SUITE_P(smoke_ScaledAttn_GPU,
                          ScaledAttnLayerGPUTest,
                          params,
                          ScaledAttnLayerGPUTest::getTestCaseName);
 
+const std::vector<std::vector<InputShape>> static_shapes{
+    // static shapes
+    {
+        // q shape
+        {ov::test::InputShape{ov::PartialShape{1, 8, 100, 128},
+            {ov::Shape{1, 8, 100, 128}}}
+        },
+        // kv shape
+        {ov::test::InputShape{ov::PartialShape{1, 8, 100, 128},
+            {ov::Shape{1, 8, 100, 128}}}
+        },
+        // attn shape: [B, 1, -1, L0+L1]
+        {ov::test::InputShape{ov::PartialShape{1, 1, 100, 100},
+            {ov::Shape{1, 1, 100, 100}}}
+        },
+    },
+    {
+        // q shape
+        {ov::test::InputShape{ov::PartialShape{1, 8, 64, 128},
+            {ov::Shape{1, 8, 64, 128}}}
+        },
+        // kv shape
+        {ov::test::InputShape{ov::PartialShape{1, 8, 13, 128},
+            {ov::Shape{1, 8, 13, 128}}}
+        },
+        // attn shape: [B, 1, -1, L0+L1]
+        {ov::test::InputShape{ov::PartialShape{1, 1, 64, 13},
+            {ov::Shape{1, 1, 64, 13}}}
+        },
+    },
+};
+
+const auto static_params = testing::Combine(testing::Values(ov::element::f16),
+                                                            testing::ValuesIn(static_shapes),
+                                                            testing::Values(true, false),
+                                                            testing::Values(true, false),
+                                                            testing::Values(true, false),
+                                                            testing::ValuesIn({disable_transpose, transpose_all}));
+
+INSTANTIATE_TEST_SUITE_P(smoke_ScaledAttnStatic_GPU,
+                         ScaledAttnLayerGPUTest,
+                         static_params,
+                         ScaledAttnLayerGPUTest::getTestCaseName);
+
 } // namespace