diff --git a/core/framework/kv_cache/kv_cache_layerwise.cpp b/core/framework/kv_cache/kv_cache_layerwise.cpp
index 9ec86f81..189221d5 100644
--- a/core/framework/kv_cache/kv_cache_layerwise.cpp
+++ b/core/framework/kv_cache/kv_cache_layerwise.cpp
@@ -28,6 +28,8 @@ limitations under the License.
 // a block will have idle lanes — we pad head_dim to the next multiple of
 // 64 on ILU to avoid this.
 
+#include "framework/kv_cache/kv_cache_layerwise.h"
+
 #include "framework/kv_cache/kv_cache.h"
 
 #include <glog/logging.h>
diff --git a/tests/core/test_layerwise_split_kv_cache.cpp b/tests/core/test_layerwise_split_kv_cache.cpp
index a2e0ce22..33c17f2b 100644
--- a/tests/core/test_layerwise_split_kv_cache.cpp
+++ b/tests/core/test_layerwise_split_kv_cache.cpp
@@ -30,12 +30,12 @@ limitations under the License.
 #include <unordered_set>
 #include <vector>
 
-#include "core/config/parallel_config_layerwise.h"
-#include "core/distributed_runtime/layerwise_split_engine_ext.h"
-#include "core/distributed_runtime/layerwise_split_master.h"
-#include "core/framework/kv_cache/kv_cache_estimation_layerwise.h"
-#include "core/framework/kv_cache/layerwise_split_layout.h"
-#include "core/framework/parallel_state/mapping_ilu.h"
+#include "config/parallel_config_layerwise.h"
+#include "distributed_runtime/layerwise_split_engine_ext.h"
+#include "distributed_runtime/layerwise_split_master.h"
+#include "framework/kv_cache/kv_cache_estimation_layerwise.h"
+#include "framework/kv_cache/layerwise_split_layout.h"
+#include "framework/parallel_state/mapping_ilu.h"
 
 namespace xllm {
 namespace {