diff --git a/paddle/fluid/inference/tensorrt/convert/cross_multihead_matmul_op.cc b/paddle/fluid/inference/tensorrt/convert/cross_multihead_matmul_op.cc
index df5665b75b34e..33c151a24f7b2 100644
--- a/paddle/fluid/inference/tensorrt/convert/cross_multihead_matmul_op.cc
+++ b/paddle/fluid/inference/tensorrt/convert/cross_multihead_matmul_op.cc
@@ -234,16 +234,13 @@ class CrossMultiheadMatMulOpConverter : public OpConverter {
     auto creator = GetPluginRegistry()->getPluginCreator("fMHCA", "1");
     assert(creator != nullptr);
     std::vector<nvinfer1::PluginField> fields{};
-    nvinfer1::PluginFieldCollection* plugin_collection =
-        static_cast<nvinfer1::PluginFieldCollection*>(
-            malloc(sizeof(*plugin_collection) +
-                   fields.size() *
-                       sizeof(nvinfer1::PluginField)));  // remember to free
+    std::unique_ptr<nvinfer1::PluginFieldCollection> plugin_collection(
+        new nvinfer1::PluginFieldCollection);
 
     plugin_collection->nbFields = static_cast<int>(fields.size());
     plugin_collection->fields = fields.data();
-    auto plugin = creator->createPlugin("fMHA_V2", plugin_collection);
-    free(plugin_collection);
+    auto plugin = creator->createPlugin("fMHA_V2", plugin_collection.get());
+    plugin_collection.reset();
     std::vector<nvinfer1::ITensor*> plugin_inputs;
     plugin_inputs.emplace_back(reshape_after_fc_q_layer->getOutput(0));
     plugin_inputs.emplace_back(reshape_after_fc_layer->getOutput(0));
diff --git a/paddle/fluid/inference/tensorrt/convert/emb_eltwise_layernorm.cc b/paddle/fluid/inference/tensorrt/convert/emb_eltwise_layernorm.cc
index 340f16330a2e5..71159e3009b6c 100644
--- a/paddle/fluid/inference/tensorrt/convert/emb_eltwise_layernorm.cc
+++ b/paddle/fluid/inference/tensorrt/convert/emb_eltwise_layernorm.cc
@@ -119,10 +119,8 @@ class EmbEltwiseLayerNormOpConverter : public OpConverter {
                             static_cast<int32_t>(emb_sizes[i]));
       }
 
-      nvinfer1::PluginFieldCollection* plugin_ptr =
-          static_cast<nvinfer1::PluginFieldCollection*>(
-              malloc(sizeof(*plugin_ptr) +
-                     fields.size() * sizeof(nvinfer1::PluginField)));
+      std::unique_ptr<nvinfer1::PluginFieldCollection> plugin_ptr(
+          new nvinfer1::PluginFieldCollection);
       plugin_ptr->nbFields = static_cast<int>(fields.size());
       plugin_ptr->fields = fields.data();
 
@@ -132,7 +130,7 @@ class EmbEltwiseLayerNormOpConverter : public OpConverter {
       auto creator = GetPluginRegistry()->getPluginCreator(
           "ManyEmbLayerNormVarlenPluginDynamic", "1");
       auto plugin_obj = creator->createPlugin(
-          "ManyEmbLayerNormVarlenPluginDynamic", plugin_ptr);
+          "ManyEmbLayerNormVarlenPluginDynamic", plugin_ptr.get());
 
       auto plugin_layer = engine_->network()->addPluginV2(
           plugin_inputs.data(), plugin_inputs.size(), *plugin_obj);
@@ -140,7 +138,7 @@ class EmbEltwiseLayerNormOpConverter : public OpConverter {
       plugin_layer->setName(("ManyEmbLayerNormVarlenPluginDynamicV1(Output: " +
                              op_desc.Output("Out")[0] + ")")
                                 .c_str());
-      free(plugin_ptr);
+      plugin_ptr.reset();
       if (enable_int8) {
         float out_scale =
             PADDLE_GET_CONST(float, op_desc.GetAttr("out_threshold"));
@@ -218,10 +216,8 @@ class EmbEltwiseLayerNormOpConverter : public OpConverter {
                             static_cast<int32_t>(emb_sizes[i]));
       }
 
-      nvinfer1::PluginFieldCollection* plugin_ptr =
-          static_cast<nvinfer1::PluginFieldCollection*>(
-              malloc(sizeof(*plugin_ptr) +
-                     fields.size() * sizeof(nvinfer1::PluginField)));
+      std::unique_ptr<nvinfer1::PluginFieldCollection> plugin_ptr(
+          new nvinfer1::PluginFieldCollection);
       plugin_ptr->nbFields = static_cast<int>(fields.size());
       plugin_ptr->fields = fields.data();
 
@@ -229,8 +225,8 @@ class EmbEltwiseLayerNormOpConverter : public OpConverter {
 
       auto creator = GetPluginRegistry()->getPluginCreator(
           "ManyEmbLayerNormPluginDynamic", "1");
-      auto plugin_obj =
-          creator->createPlugin("ManyEmbLayerNormPluginDynamic", plugin_ptr);
+      auto plugin_obj = creator->createPlugin("ManyEmbLayerNormPluginDynamic",
+                                              plugin_ptr.get());
 
       auto plugin_layer = engine_->network()->addPluginV2(
           plugin_inputs.data(), plugin_inputs.size(), *plugin_obj);
@@ -238,7 +234,7 @@ class EmbEltwiseLayerNormOpConverter : public OpConverter {
       plugin_layer->setName(("ManyEmbLayerNormPluginDynamicV1(Output: " +
                              op_desc.Output("Out")[0] + ")")
                                 .c_str());
-      free(plugin_ptr);
+      plugin_ptr.reset();
       if (enable_int8) {
         float out_scale =
             PADDLE_GET_CONST(float, op_desc.GetAttr("out_threshold"));
diff --git a/paddle/fluid/inference/tensorrt/convert/flash_multihead_matmul_op.cc b/paddle/fluid/inference/tensorrt/convert/flash_multihead_matmul_op.cc
index e5904a1cf7543..757f90181cd0d 100644
--- a/paddle/fluid/inference/tensorrt/convert/flash_multihead_matmul_op.cc
+++ b/paddle/fluid/inference/tensorrt/convert/flash_multihead_matmul_op.cc
@@ -210,16 +210,12 @@ class FlashMultiheadMatMulOpConverter : public OpConverter {
     auto creator = GetPluginRegistry()->getPluginCreator("fMHA_V2", "1");
     assert("fmha_v2 plugin creater must not be null" && creator != nullptr);
     std::vector<nvinfer1::PluginField> fields{};
-    nvinfer1::PluginFieldCollection* plugin_collection =
-        static_cast<nvinfer1::PluginFieldCollection*>(
-            malloc(sizeof(*plugin_collection) +
-                   fields.size() *
-                       sizeof(nvinfer1::PluginField)));  // remember to free
-
+    std::unique_ptr<nvinfer1::PluginFieldCollection> plugin_collection(
+        new nvinfer1::PluginFieldCollection);
     plugin_collection->nbFields = static_cast<int>(fields.size());
     plugin_collection->fields = fields.data();
-    auto plugin = creator->createPlugin("fMHA_V2", plugin_collection);
-    free(plugin_collection);
+    auto plugin = creator->createPlugin("fMHA_V2", plugin_collection.get());
+    plugin_collection.reset();
     std::vector<nvinfer1::ITensor*> plugin_inputs;
     plugin_inputs.emplace_back(reshape_before_mha_layer->getOutput(0));
     auto plugin_layer = engine_->network()->addPluginV2(
diff --git a/paddle/fluid/inference/tensorrt/convert/multiclass_nms3_op.cc b/paddle/fluid/inference/tensorrt/convert/multiclass_nms3_op.cc
index 449593533820b..107217477d14f 100644
--- a/paddle/fluid/inference/tensorrt/convert/multiclass_nms3_op.cc
+++ b/paddle/fluid/inference/tensorrt/convert/multiclass_nms3_op.cc
@@ -102,10 +102,8 @@ class MultiClassNMS3OpConverter : public OpConverter {
         {"clipBoxes", &clip_boxes, nvinfer1::PluginFieldType::kINT32, 1},
     };
 
-    nvinfer1::PluginFieldCollection* plugin_collections =
-        static_cast<nvinfer1::PluginFieldCollection*>(
-            malloc(sizeof(*plugin_collections) +
-                   fields.size() * sizeof(nvinfer1::PluginField)));
+    std::unique_ptr<nvinfer1::PluginFieldCollection> plugin_collections(
+        new nvinfer1::PluginFieldCollection);
     plugin_collections->nbFields = static_cast<int>(fields.size());
     plugin_collections->fields = fields.data();
     std::string nms_plugin_name = "BatchedNMS_TRT";
@@ -114,9 +112,9 @@ class MultiClassNMS3OpConverter : public OpConverter {
     }
     auto creator =
         GetPluginRegistry()->getPluginCreator(nms_plugin_name.c_str(), "1");
-    auto batch_nms_plugin =
-        creator->createPlugin(nms_plugin_name.c_str(), plugin_collections);
-    free(plugin_collections);
+    auto batch_nms_plugin = creator->createPlugin(nms_plugin_name.c_str(),
+                                                  plugin_collections.get());
+    plugin_collections.reset();
 
     auto batch_nms_layer = engine_->network()->addPluginV2(
         batch_nms_inputs.data(), batch_nms_inputs.size(), *batch_nms_plugin);
diff --git a/paddle/fluid/inference/tensorrt/convert/multiclass_nms_op.cc b/paddle/fluid/inference/tensorrt/convert/multiclass_nms_op.cc
index e14ee099aa0f8..ab0dacedd2d07 100644
--- a/paddle/fluid/inference/tensorrt/convert/multiclass_nms_op.cc
+++ b/paddle/fluid/inference/tensorrt/convert/multiclass_nms_op.cc
@@ -100,10 +100,8 @@ class MultiClassNMSOpConverter : public OpConverter {
         {"clipBoxes", &clip_boxes, nvinfer1::PluginFieldType::kINT32, 1},
     };
 
-    nvinfer1::PluginFieldCollection* plugin_collections =
-        static_cast<nvinfer1::PluginFieldCollection*>(
-            malloc(sizeof(*plugin_collections) +
-                   fields.size() * sizeof(nvinfer1::PluginField)));
+    std::unique_ptr<nvinfer1::PluginFieldCollection> plugin_collections(
+        new nvinfer1::PluginFieldCollection);
     plugin_collections->nbFields = static_cast<int>(fields.size());
     plugin_collections->fields = fields.data();
 
@@ -113,9 +111,9 @@ class MultiClassNMSOpConverter : public OpConverter {
     }
     auto creator =
         GetPluginRegistry()->getPluginCreator(nms_plugin_name.c_str(), "1");
-    auto batch_nms_plugin =
-        creator->createPlugin(nms_plugin_name.c_str(), plugin_collections);
-    free(plugin_collections);
+    auto batch_nms_plugin = creator->createPlugin(nms_plugin_name.c_str(),
+                                                  plugin_collections.get());
+    plugin_collections.reset();
 
     auto batch_nms_layer = engine_->network()->addPluginV2(
         batch_nms_inputs.data(), batch_nms_inputs.size(), *batch_nms_plugin);
diff --git a/paddle/fluid/inference/tensorrt/convert/multihead_matmul_op.cc b/paddle/fluid/inference/tensorrt/convert/multihead_matmul_op.cc
index 73c43d39357c0..a092b3215502e 100644
--- a/paddle/fluid/inference/tensorrt/convert/multihead_matmul_op.cc
+++ b/paddle/fluid/inference/tensorrt/convert/multihead_matmul_op.cc
@@ -234,17 +234,14 @@ class MultiheadMatMulOpConverter : public OpConverter {
                               nvinfer1::PluginFieldType::kFLOAT32,
                               1});
           }
-          nvinfer1::PluginFieldCollection* plugin_collection =
-              static_cast<nvinfer1::PluginFieldCollection*>(malloc(
-                  sizeof(*plugin_collection) +
-                  fields.size() *
-                      sizeof(nvinfer1::PluginField)));  // remember to free
+          std::unique_ptr<nvinfer1::PluginFieldCollection> plugin_collection(
+              new nvinfer1::PluginFieldCollection);
           plugin_collection->nbFields = static_cast<int>(fields.size());
           plugin_collection->fields = fields.data();
 
           auto plugin = creator->createPlugin("CustomQKVToContextPluginDynamic",
-                                              plugin_collection);
-          free(plugin_collection);
+                                              plugin_collection.get());
+          plugin_collection.reset();
 
           std::vector<nvinfer1::ITensor*> plugin_inputs;
           plugin_inputs.emplace_back(fc_layer->getOutput(0));
@@ -429,17 +426,14 @@ class MultiheadMatMulOpConverter : public OpConverter {
                               nvinfer1::PluginFieldType::kFLOAT32,
                               1});
           }
-          nvinfer1::PluginFieldCollection* plugin_collection =
-              static_cast<nvinfer1::PluginFieldCollection*>(malloc(
-                  sizeof(*plugin_collection) +
-                  fields.size() *
-                      sizeof(nvinfer1::PluginField)));  // remember to free
+          std::unique_ptr<nvinfer1::PluginFieldCollection> plugin_collection(
+              new nvinfer1::PluginFieldCollection);
           plugin_collection->nbFields = static_cast<int>(fields.size());
           plugin_collection->fields = fields.data();
 
           auto plugin = creator->createPlugin("CustomQKVToContextPluginDynamic",
-                                              plugin_collection);
-          free(plugin_collection);
+                                              plugin_collection.get());
+          plugin_collection.reset();
 
           std::vector<nvinfer1::ITensor*> plugin_inputs;
           plugin_inputs.emplace_back(
@@ -661,16 +655,13 @@ class MultiheadMatMulOpConverter : public OpConverter {
                &var_seqlen,
                nvinfer1::PluginFieldType::kINT32,
                1}};
-          nvinfer1::PluginFieldCollection* plugin_collection =
-              static_cast<nvinfer1::PluginFieldCollection*>(malloc(
-                  sizeof(*plugin_collection) +
-                  fields.size() *
-                      sizeof(nvinfer1::PluginField)));  // remember to free
+          std::unique_ptr<nvinfer1::PluginFieldCollection> plugin_collection(
+              new nvinfer1::PluginFieldCollection);
           plugin_collection->nbFields = static_cast<int>(fields.size());
           plugin_collection->fields = fields.data();
           auto plugin = creator->createPlugin("CustomQKVToContextPluginDynamic",
-                                              plugin_collection);
-          free(plugin_collection);
+                                              plugin_collection.get());
+          plugin_collection.reset();
           // set inputs
           std::vector<nvinfer1::ITensor*> plugin_inputs;
           // input_0 for plugin
diff --git a/paddle/fluid/inference/tensorrt/convert/preln_emb_eltwise_layernorm.cc b/paddle/fluid/inference/tensorrt/convert/preln_emb_eltwise_layernorm.cc
index 0ec1336f0e2d1..9f9cbe7c6bceb 100644
--- a/paddle/fluid/inference/tensorrt/convert/preln_emb_eltwise_layernorm.cc
+++ b/paddle/fluid/inference/tensorrt/convert/preln_emb_eltwise_layernorm.cc
@@ -173,10 +173,8 @@ class PrelnEmbEltwiseLayerNormOpConverter : public OpConverter {
                           static_cast<int32_t>(emb_sizes[i]));
     }
 
-    nvinfer1::PluginFieldCollection* plugin_ptr =
-        static_cast<nvinfer1::PluginFieldCollection*>(
-            malloc(sizeof(*plugin_ptr) +
-                   fields.size() * sizeof(nvinfer1::PluginField)));
+    std::unique_ptr<nvinfer1::PluginFieldCollection> plugin_ptr(
+        new nvinfer1::PluginFieldCollection);
     plugin_ptr->nbFields = static_cast<int>(fields.size());
     plugin_ptr->fields = fields.data();
 
@@ -188,7 +186,7 @@ class PrelnEmbEltwiseLayerNormOpConverter : public OpConverter {
         "ManyEmbLayerNormVarlenPluginDynamic", "2");
 
     auto plugin_obj = creator->createPlugin(
-        "ManyEmbLayerNormVarlenPluginDynamic", plugin_ptr);
+        "ManyEmbLayerNormVarlenPluginDynamic", plugin_ptr.get());
 
     auto plugin_layer = engine_->network()->addPluginV2(
         plugin_inputs.data(), plugin_inputs.size(), *plugin_obj);
@@ -196,7 +194,7 @@ class PrelnEmbEltwiseLayerNormOpConverter : public OpConverter {
     plugin_layer->setName(("ManyEmbLayerNormPluginDynamic_V3(Output: " +
                            op_desc.Output("Out")[0] + ")")
                               .c_str());
-    free(plugin_ptr);
+    plugin_ptr.reset();
     float out_0_scale =
         PADDLE_GET_CONST(float, op_desc.GetAttr("out_0_threshold"));
     float out_1_scale =
diff --git a/paddle/fluid/inference/tensorrt/convert/prompt_tuning_emb_eltwise_layernorm.cc b/paddle/fluid/inference/tensorrt/convert/prompt_tuning_emb_eltwise_layernorm.cc
index 9250807662543..e6beaae910d96 100644
--- a/paddle/fluid/inference/tensorrt/convert/prompt_tuning_emb_eltwise_layernorm.cc
+++ b/paddle/fluid/inference/tensorrt/convert/prompt_tuning_emb_eltwise_layernorm.cc
@@ -118,10 +118,8 @@ class PromptTuningEmbEltwiseLayerNormOpConverter : public OpConverter {
                           static_cast<int32_t>(emb_sizes[i]));
     }
 
-    nvinfer1::PluginFieldCollection* plugin_ptr =
-        static_cast<nvinfer1::PluginFieldCollection*>(
-            malloc(sizeof(*plugin_ptr) +
-                   fields.size() * sizeof(nvinfer1::PluginField)));
+    std::unique_ptr<nvinfer1::PluginFieldCollection> plugin_ptr(
+        new nvinfer1::PluginFieldCollection);
     plugin_ptr->nbFields = static_cast<int>(fields.size());
     plugin_ptr->fields = fields.data();
 
@@ -134,7 +132,7 @@ class PromptTuningEmbEltwiseLayerNormOpConverter : public OpConverter {
     auto creator = GetPluginRegistry()->getPluginCreator(
         "PromptTuningEmbLayerNormVarlenPluginDynamic", "1");
     auto plugin_obj = creator->createPlugin(
-        "PromptTuningEmbLayerNormVarlenPluginDynamic", plugin_ptr);
+        "PromptTuningEmbLayerNormVarlenPluginDynamic", plugin_ptr.get());
 
     auto plugin_layer = engine_->network()->addPluginV2(
         plugin_inputs.data(), plugin_inputs.size(), *plugin_obj);
@@ -143,7 +141,7 @@ class PromptTuningEmbEltwiseLayerNormOpConverter : public OpConverter {
         ("PromptTuningEmbLayerNormVarlenPluginDynamicV1(Output: " +
          op_desc.Output("Out")[0] + ")")
             .c_str());
-    free(plugin_ptr);
+    plugin_ptr.reset();
     if (enable_int8) {
       float out_scale =
           PADDLE_GET_CONST(float, op_desc.GetAttr("out_threshold"));
diff --git a/paddle/fluid/inference/tensorrt/convert/qk_multihead_matmul_op.cc b/paddle/fluid/inference/tensorrt/convert/qk_multihead_matmul_op.cc
index e8ed4af9cddf7..aafbec6660c67 100644
--- a/paddle/fluid/inference/tensorrt/convert/qk_multihead_matmul_op.cc
+++ b/paddle/fluid/inference/tensorrt/convert/qk_multihead_matmul_op.cc
@@ -258,16 +258,12 @@ class QkMultiheadMatMulOpConverter : public OpConverter {
     auto creator = GetPluginRegistry()->getPluginCreator("fMHA_V2", "1");
     assert(creator != nullptr);
     std::vector<nvinfer1::PluginField> fields{};
-    nvinfer1::PluginFieldCollection* plugin_collection =
-        static_cast<nvinfer1::PluginFieldCollection*>(
-            malloc(sizeof(*plugin_collection) +
-                   fields.size() *
-                       sizeof(nvinfer1::PluginField)));  // remember to free
-
+    std::unique_ptr<nvinfer1::PluginFieldCollection> plugin_collection(
+        new nvinfer1::PluginFieldCollection);
     plugin_collection->nbFields = static_cast<int>(fields.size());
     plugin_collection->fields = fields.data();
-    auto plugin = creator->createPlugin("fMHA_V2", plugin_collection);
-    free(plugin_collection);
+    auto plugin = creator->createPlugin("fMHA_V2", plugin_collection.get());
+    plugin_collection.reset();
     std::vector<nvinfer1::ITensor*> plugin_inputs;
     plugin_inputs.emplace_back(mha_input_tensor);
     auto plugin_layer = engine_->network()->addPluginV2(
diff --git a/paddle/fluid/inference/tensorrt/convert/skip_layernorm.cc b/paddle/fluid/inference/tensorrt/convert/skip_layernorm.cc
index ab70ebb6ccd81..988d0d064c862 100644
--- a/paddle/fluid/inference/tensorrt/convert/skip_layernorm.cc
+++ b/paddle/fluid/inference/tensorrt/convert/skip_layernorm.cc
@@ -149,17 +149,15 @@ class SkipLayerNormOpConverter : public OpConverter {
            scale_weight.values,
            GetPluginFieldType(scale_weight.type),
            static_cast<int32_t>(scale_weight.count)}};
-      nvinfer1::PluginFieldCollection* pluginPtr =
-          static_cast<nvinfer1::PluginFieldCollection*>(
-              malloc(sizeof(nvinfer1::PluginFieldCollection) +
-                     fields.size() * sizeof(nvinfer1::PluginField)));
+      std::unique_ptr<nvinfer1::PluginFieldCollection> pluginPtr(
+          new nvinfer1::PluginFieldCollection);
       pluginPtr->nbFields = static_cast<int32_t>(fields.size());
       pluginPtr->fields = fields.data();
 
-      auto pluginObj =
-          creator->createPlugin("CustomSkipLayerNormPluginDynamic", pluginPtr);
+      auto pluginObj = creator->createPlugin("CustomSkipLayerNormPluginDynamic",
+                                             pluginPtr.get());
 
-      free(pluginPtr);
+      pluginPtr.reset();
 
       auto plugin_layer = engine_->network()->addPluginV2(
           inputs.data(), inputs.size(), *pluginObj);
@@ -213,18 +211,15 @@ class SkipLayerNormOpConverter : public OpConverter {
                           smooth_scale.data(),
                           nvinfer1::PluginFieldType::kFLOAT32,
                           static_cast<int32_t>(smooth_scale.size())});
-        nvinfer1::PluginFieldCollection* pluginPtr =
-            static_cast<nvinfer1::PluginFieldCollection*>(
-                malloc(sizeof(nvinfer1::PluginFieldCollection) +
-                       fields.size() *
-                           sizeof(nvinfer1::PluginField)));  // remember to free
+        std::unique_ptr<nvinfer1::PluginFieldCollection> pluginPtr(
+            new nvinfer1::PluginFieldCollection);
         pluginPtr->nbFields = static_cast<int32_t>(fields.size());
         pluginPtr->fields = fields.data();
 
         auto pluginObj = creator->createPlugin(
-            "CustomSkipLayerNormPluginDynamicWithSmooth", pluginPtr);
+            "CustomSkipLayerNormPluginDynamicWithSmooth", pluginPtr.get());
 
-        free(pluginPtr);
+        pluginPtr.reset();
 
         auto plugin_layer = engine_->network()->addPluginV2(
             inputs.data(), inputs.size(), *pluginObj);
@@ -237,18 +232,15 @@ class SkipLayerNormOpConverter : public OpConverter {
                 "layer"));
         layer = plugin_layer;
       } else {
-        nvinfer1::PluginFieldCollection* pluginPtr =
-            static_cast<nvinfer1::PluginFieldCollection*>(
-                malloc(sizeof(nvinfer1::PluginFieldCollection) +
-                       fields.size() *
-                           sizeof(nvinfer1::PluginField)));  // remember to free
+        std::unique_ptr<nvinfer1::PluginFieldCollection> pluginPtr(
+            new nvinfer1::PluginFieldCollection);
         pluginPtr->nbFields = static_cast<int32_t>(fields.size());
         pluginPtr->fields = fields.data();
 
         auto pluginObj = creator->createPlugin(
-            "CustomSkipLayerNormPluginDynamic", pluginPtr);
+            "CustomSkipLayerNormPluginDynamic", pluginPtr.get());
 
-        free(pluginPtr);
+        pluginPtr.reset();
 
         auto plugin_layer = engine_->network()->addPluginV2(
             inputs.data(), inputs.size(), *pluginObj);
diff --git a/paddle/pir/include/core/interface_value.h b/paddle/pir/include/core/interface_value.h
index 64619a0e0f591..030a6065b8415 100644
--- a/paddle/pir/include/core/interface_value.h
+++ b/paddle/pir/include/core/interface_value.h
@@ -14,6 +14,7 @@
 
 #pragma once
 
+#include <memory>
 #include <set>
 #include <type_traits>
 
@@ -27,7 +28,7 @@ class IR_API InterfaceValue {
   template <typename Interface, typename Model>
   static InterfaceValue Get();
   TypeId type_id() const { return type_id_; }
-  void *model() const { return model_; }
+  void *model() const { return model_.get(); }
 
   InterfaceValue() = default;
   InterfaceValue(TypeId type_id) : type_id_(type_id) {}  // NOLINT
@@ -35,7 +36,7 @@ class IR_API InterfaceValue {
   InterfaceValue(InterfaceValue &&) noexcept;
   InterfaceValue &operator=(const InterfaceValue &) = delete;
   InterfaceValue &operator=(InterfaceValue &&) noexcept;
-  ~InterfaceValue();
+  ~InterfaceValue() = default;
   void swap(InterfaceValue &&val) {
     using std::swap;
     swap(type_id_, val.type_id_);
@@ -51,7 +52,8 @@ class IR_API InterfaceValue {
 
  private:
   TypeId type_id_;
-  void *model_{nullptr};
+  std::unique_ptr<void, decltype(static_cast<void (*)(void *)>(free))> model_{
+      nullptr, static_cast<void (*)(void *)>(free)};
 };
 
 template <typename Interface, typename Model>
@@ -64,13 +66,14 @@ InterfaceValue InterfaceValue::Get() {
       sizeof(typename Interface::Concept) == sizeof(Model),
       "Compared with Concept, Model class shouldn't define new data members");
 
-  val.model_ = malloc(sizeof(Model));
-  if (val.model_ == nullptr) {
+  void *model_raw = malloc(sizeof(Model));
+  if (model_raw == nullptr) {
     throw("Alloc memory for interface failed.");
   }
   static_assert(std::is_trivially_destructible<Model>::value,
                 "interface models must be trivially destructible");
-  new (val.model_) Model();
+  new (model_raw) Model();
+  val.model_.reset(model_raw);
   return val;
 }
 
diff --git a/paddle/pir/src/core/interface_support.cc b/paddle/pir/src/core/interface_support.cc
index 973b98668a061..0d2a87fc6bd66 100644
--- a/paddle/pir/src/core/interface_support.cc
+++ b/paddle/pir/src/core/interface_support.cc
@@ -15,14 +15,9 @@
 #include "paddle/pir/include/core/interface_support.h"
 
 namespace pir {
-InterfaceValue::~InterfaceValue() {
-  if (model_) free(model_);
-}
-
 InterfaceValue::InterfaceValue(InterfaceValue&& val) noexcept {
   type_id_ = val.type_id_;
-  model_ = val.model_;
-  val.model_ = nullptr;
+  model_ = std::move(val.model_);
 }
 
 InterfaceValue& InterfaceValue::operator=(InterfaceValue&& val) noexcept {
diff --git a/test/deprecated/cpp/inference/api/analyzer_capi_exp_pd_threads_tester.cc b/test/deprecated/cpp/inference/api/analyzer_capi_exp_pd_threads_tester.cc
index b06c637c86e47..1b66f7c246f7e 100644
--- a/test/deprecated/cpp/inference/api/analyzer_capi_exp_pd_threads_tester.cc
+++ b/test/deprecated/cpp/inference/api/analyzer_capi_exp_pd_threads_tester.cc
@@ -39,7 +39,7 @@ typedef struct RunParameter {
   size_t shape_size;
   float* input_data;
   int32_t out_size;
-  float* out_data;
+  std::vector<float> out_data;
   int32_t thread_index;
 } RunParameter;
 
@@ -62,9 +62,8 @@ void* run(void* thread_param) {
     param->out_size = param->out_size * output_shape->data[index];
   }
   PD_OneDimArrayInt32Destroy(output_shape);
-  param->out_data =
-      reinterpret_cast<float*>(malloc(param->out_size * sizeof(float)));
-  PD_TensorCopyToCpuFloat(output_tensor, param->out_data);
+  param->out_data.resize(param->out_size);
+  PD_TensorCopyToCpuFloat(output_tensor, param->out_data.data());
   PD_TensorDestroy(output_tensor);
   PD_OneDimArrayCstrDestroy(output_names);
   PD_TensorDestroy(tensor);
@@ -80,23 +79,19 @@ void threads_run(int thread_num) {
                     (model_dir + "/__params__").c_str());
   PD_Predictor* predictor = PD_PredictorCreate(config);
 
-  pthread_t* threads =
-      reinterpret_cast<pthread_t*>(malloc(thread_num * sizeof(pthread_t)));
-  RunParameter* params = reinterpret_cast<RunParameter*>(
-      malloc(thread_num * sizeof(RunParameter)));
+  std::vector<pthread_t> threads(thread_num);
+  std::vector<RunParameter> params(thread_num);
+
   std::array<int32_t, 4> shapes = {1, 3, 300, 300};
-  float* input =
-      reinterpret_cast<float*>(malloc(1 * 3 * 300 * 300 * sizeof(float)));
-  memset(input, 0, 1 * 3 * 300 * 300 * sizeof(float));
+  std::vector<float> input(1 * 3 * 300 * 300, 0);
   for (int i = 0; i < thread_num; ++i) {
     params[i].predictor = PD_PredictorClone(predictor);
     params[i].shapes = shapes.data();
     params[i].shape_size = 4;
-    params[i].input_data = input;
+    params[i].input_data = input.data();
     params[i].out_size = 0;
-    params[i].out_data = nullptr;
     params[i].thread_index = i;
-    pthread_create(&(threads[i]), nullptr, run, (params + i));
+    pthread_create(&(threads[i]), nullptr, run, &(params[i]));
   }
   for (int i = 0; i < thread_num; ++i) {
     pthread_join(threads[i], nullptr);
@@ -111,11 +106,7 @@ void threads_run(int thread_num) {
   }
   for (int i = 0; i < thread_num; ++i) {
     PD_PredictorDestroy(params[i].predictor);
-    free(params[i].out_data);
   }
-  free(input);
-  free(params);
-  free(threads);
   PD_PredictorDestroy(predictor);
 }