diff --git a/paddle/fluid/framework/naive_executor.cc b/paddle/fluid/framework/naive_executor.cc
index 556a167e74f226bc838b020809ada790f464bfdd..c31e0661140daf1c99c82f4465efcb9769577178 100644
--- a/paddle/fluid/framework/naive_executor.cc
+++ b/paddle/fluid/framework/naive_executor.cc
@@ -66,13 +66,6 @@ void NaiveExecutor::Run() {
                                 platform::NvtxRangeColor::Green);
 #endif
 
-    // According to reuse table, we share the out tensor's holder.
-    if (reuse_cache_.count(op.get())) {
-      for (auto &it : reuse_cache_[op.get()]) {
-        it.first->ShareBufferWith(*cluster_buffer_[it.second], true);
-      }
-    }
-
     op->Run(*scope_, place_);
 
     // Update the shared_holder so that only records the max one.
@@ -81,6 +74,22 @@ void NaiveExecutor::Run() {
         if (it.first->memory_size() >
             cluster_buffer_[it.second]->memory_size()) {
           cluster_buffer_[it.second] = it.first;
+          int updated_cluster_id = it.second;
+
+          // cluster_buffer_[it.second] has been updated to be a new
+          // phi::DenseTensor*, we need change all phi::DenseTensor's
+          // shared_holder in this cluster. The following two loops code looks
+          // ugly, it does work. The following two loops seem time-consuming,
+          // but once the memory reaches its peak, the cluster will not update,
+          // so it's ok.
+          for (auto &op_map : reuse_cache_) {
+            // op_map.second is std::unordered_map<phi::DenseTensor*, int>.
+            for (auto &it2 : op_map.second) {
+              if (it2.second == updated_cluster_id) {
+                it2.first->ShareBufferWith(*cluster_buffer_[it2.second], true);
+              }
+            }
+          }
         }
       }
     }