[GPU] GQA optimization of PagedAttention OCL kernel for long sequences

Author: sshlyapn

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

@@ -660,7 +660,7 @@ struct paged_attention_impl : multi_stage_primitive<paged_attention> {
 
         if (desc->heads_num != desc->kv_heads_num) {
             config.broadcast_axis = 1;
-            config.group_size = desc->heads_num / desc->kv_heads_num;
+            config.kv_group_size = desc->heads_num / desc->kv_heads_num;
         }
 
         if (desc->has_scores_output() && !is_dynamic) {

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

@@ -252,7 +252,7 @@ struct scaled_dot_product_attention_impl : multi_stage_primitive<scaled_dot_prod
         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();
+                config.kv_group_size = query_shape[num_heads_dim].get_length() / key_shape[num_heads_dim].get_length();
             }
         }