diff --git a/paddle/math/tests/CMakeLists.txt b/paddle/math/tests/CMakeLists.txt
index ceb96b2e250d8e04ffb2b1d8c77ad498dca91cf3..bdecba0869d9cda47cf7bcd03f898804fdb9a829 100644
--- a/paddle/math/tests/CMakeLists.txt
+++ b/paddle/math/tests/CMakeLists.txt
@@ -31,3 +31,4 @@ add_simple_unittest(test_FPException)
 add_simple_unittest(test_GpuProfiler)
 add_simple_unittest(test_BaseMatrix)
 add_simple_unittest(test_Matrix)
+add_simple_unittest(test_Matrix2)
diff --git a/paddle/optimizer/Tensor.h b/paddle/optimizer/Tensor.h
index 2eefdec3e5e3410a60bc1f57761930af4e6c8a45..49b8dadaaef219517b15cfc99fc47f4eaabf9b83 100644
--- a/paddle/optimizer/Tensor.h
+++ b/paddle/optimizer/Tensor.h
@@ -5,6 +5,7 @@
 
 #include <string.h>
 #include <memory>
+#include "paddle/math/MemoryHandle.h"
 #include "paddle/utils/Common.h"
 #include "paddle/utils/Logging.h"
 
@@ -14,15 +15,19 @@ namespace optimizer {
 template <class T>
 class TensorT {
 public:
-  TensorT(size_t size) : height_(1), width_(size) { data_ = new T[size]; }
+  TensorT(size_t size)
+      : TensorT(std::make_shared<CpuMemoryHandle>(size * sizeof(float)), size) {
+  }
+  TensorT(CpuMemHandlePtr handle, size_t size)
+      : height_(1),
+        width_(size),
+        data_(reinterpret_cast<T*>(handle->getBuf())) {}
 
   TensorT(T* data, size_t size) : height_(1), width_(size), data_(data) {}
 
   TensorT(T* data, size_t h, size_t w) : height_(h), width_(w), data_(data) {}
 
-  ~TensorT() {
-    if (data_) delete data_;
-  }
+  virtual ~TensorT() {}
 
   T* get_buffer() { return this->data_; }