diff --git a/frameworks/native/neural_network_runtime/hdi_prepared_model_v2_0.cpp b/frameworks/native/neural_network_runtime/hdi_prepared_model_v2_0.cpp
index adca2bf15f1501a45d174d1bb278122d235078e8..b3b535b9048f43b84ee8983fc8b95dabd207038b 100644
--- a/frameworks/native/neural_network_runtime/hdi_prepared_model_v2_0.cpp
+++ b/frameworks/native/neural_network_runtime/hdi_prepared_model_v2_0.cpp
@@ -166,6 +166,17 @@ HDIPreparedModelV2_0::HDIPreparedModelV2_0(OHOS::sptr<V2_0::IPreparedModel> hdiP
     hdiPreparedModel->GetVersion(m_hdiVersion.first, m_hdiVersion.second);
 }
 
+HDIPreparedModelV2_0::~HDIPreparedModelV2_0()
+{
+    for (auto addr : m_addrs) {
+        auto memManager = MemoryManager::GetInstance();
+        OH_NN_ReturnCode ret = memManager->UnMapMemory(addr);
+        if (ret != OH_NN_SUCCESS) {
+            LOGE("~HDIPreparedModelV2_0 UnMapMemory failed.");
+        }
+    }
+}
+
 OH_NN_ReturnCode HDIPreparedModelV2_0::ExportModelCache(std::vector<Buffer>& modelCache)
 {
     if (!modelCache.empty()) {
@@ -187,6 +198,7 @@ OH_NN_ReturnCode HDIPreparedModelV2_0::ExportModelCache(std::vector<Buffer>& mod
             LOGE("Export the %{public}zuth model cache failed, cannot not map fd to address.", i + 1);
             return OH_NN_MEMORY_ERROR;
         }
+        m_addrs.emplace_back(addr);
         Buffer modelbuffer {addr, iBuffers[i].bufferSize};
         modelCache.emplace_back(modelbuffer);
     }
diff --git a/frameworks/native/neural_network_runtime/hdi_prepared_model_v2_0.h b/frameworks/native/neural_network_runtime/hdi_prepared_model_v2_0.h
index 7f3defde710e30866dd3a47b1be4d6ce63d7e917..cbbb9b9b2e19dc4d373930ca104badb2c433317d 100644
--- a/frameworks/native/neural_network_runtime/hdi_prepared_model_v2_0.h
+++ b/frameworks/native/neural_network_runtime/hdi_prepared_model_v2_0.h
@@ -34,6 +34,7 @@ namespace NeuralNetworkRuntime {
 class HDIPreparedModelV2_0 : public PreparedModel {
 public:
     explicit HDIPreparedModelV2_0(OHOS::sptr<V2_0::IPreparedModel> hdiPreparedModel);
+    ~HDIPreparedModelV2_0() override;
 
     OH_NN_ReturnCode ExportModelCache(std::vector<Buffer>& modelCache) override;
 
@@ -54,6 +55,7 @@ private:
     // first: major version, second: minor version
     std::pair<uint32_t, uint32_t> m_hdiVersion;
     OHOS::sptr<V2_0::IPreparedModel> m_hdiPreparedModel {nullptr};
+    std::vector<void *> m_addrs;
 };
 } // namespace NeuralNetworkRuntime
 } // OHOS
diff --git a/frameworks/native/neural_network_runtime/nntensor.cpp b/frameworks/native/neural_network_runtime/nntensor.cpp
index 477bdf60f84b79397d224c7bc926c62e4cb20539..d234227d863dfa96b457b482c24e37577131e18a 100644
--- a/frameworks/native/neural_network_runtime/nntensor.cpp
+++ b/frameworks/native/neural_network_runtime/nntensor.cpp
@@ -26,9 +26,7 @@ namespace OHOS {
 namespace NeuralNetworkRuntime {
 NNTensor2_0::~NNTensor2_0()
 {
-    if (!m_isUserData) {
-        ReleaseMemory();
-    }
+    ReleaseMemory();
 
     delete m_tensorDesc;
     m_tensorDesc = nullptr;
@@ -247,37 +245,39 @@ OH_NN_ReturnCode NNTensor2_0::ReleaseMemory()
         return OH_NN_INVALID_PARAMETER;
     }
 
-    BackendManager& backendManager = BackendManager::GetInstance();
-    std::shared_ptr<Backend> backend = backendManager.GetBackend(m_backendID);
-    if (backend == nullptr) {
-        LOGE("NNTensor2_0::ReleaseMemory failed, failed to get backend of %{public}zu.", m_backendID);
-        return OH_NN_NULL_PTR;
-    }
-
-    auto* nnrtBackend = reinterpret_cast<NNBackend*>(backend.get());
-    auto device = nnrtBackend->GetDevice();
-    if (device == nullptr) {
-        LOGE("");
-        return OH_NN_NULL_PTR;
-    }
-    auto oldRet = device->ReleaseBuffer(m_fd, m_size);
-    if (oldRet != OH_NN_SUCCESS) {
-        LOGE("NNTensor2_0::ReleaseMemory failed, failed to release buffer.");
-        return OH_NN_MEMORY_ERROR;
-    }
-
     auto unmapResult = munmap(m_data, m_size);
     if (unmapResult != 0) {
         LOGE("NNTensor2_0::ReleaseMemory failed. Please try again.");
         return OH_NN_MEMORY_ERROR;
+        }
+
+    if (!m_isUserData) {
+        if (close(m_fd) != 0) {
+            LOGE("NNTensor2_0::ReleaseMemory failed. fd=%{public}d", m_fd);
+            return OH_NN_MEMORY_ERROR;
+        }
+        BackendManager& backendManager = BackendManager::GetInstance();
+        std::shared_ptr<Backend> backend = backendManager.GetBackend(m_backendID);
+        if (backend == nullptr) {
+            LOGE("NNTensor2_0::ReleaseMemory failed, failed to get backend of %{public}zu.", m_backendID);
+            return OH_NN_NULL_PTR;
+        }
+
+        auto* nnrtBackend = reinterpret_cast<NNBackend*>(backend.get());
+        auto device = nnrtBackend->GetDevice();
+        if (device == nullptr) {
+            LOGE("");
+            return OH_NN_NULL_PTR;
+        }
+        auto oldRet = device->ReleaseBuffer(m_fd, m_size);
+        if (oldRet != OH_NN_SUCCESS) {
+            LOGE("NNTensor2_0::ReleaseMemory failed, failed to release buffer.");
+            return OH_NN_MEMORY_ERROR;
+        }
     }
+
     m_data = nullptr;
     m_size = 0;
-
-    if (close(m_fd) != 0) {
-        LOGE("NNTensor2_0::ReleaseMemory failed. fd=%{public}d", m_fd);
-        return OH_NN_MEMORY_ERROR;
-    }
     m_fd = 0;
 
     return OH_NN_SUCCESS;