diff --git a/paddle/infrt/kernel/phi/dense_tensor_kernels.cc b/paddle/infrt/kernel/phi/dense_tensor_kernels.cc
index 7ffc8de15107563f778ea00a1a517d13b02e0938..a9b18c769dca81f44c2f87b0eb6935b3f4da7966 100644
--- a/paddle/infrt/kernel/phi/dense_tensor_kernels.cc
+++ b/paddle/infrt/kernel/phi/dense_tensor_kernels.cc
@@ -13,7 +13,9 @@
 // limitations under the License.
 
 #include "paddle/infrt/kernel/phi/dense_tensor_kernels.h"
+#include <memory>
 #include "llvm/Support/ErrorHandling.h"
+#include "paddle/infrt/backends/host/phi_allocator.h"
 #include "paddle/infrt/common/string.h"
 #include "paddle/infrt/dialect/phi/data_type.h"
 #include "paddle/infrt/kernel/phi/context_kernels.h"
@@ -22,24 +24,13 @@
 #include "paddle/infrt/tensor/tensor_map.h"
 #include "paddle/phi/backends/all_context.h"
 #include "paddle/phi/common/place.h"
+#include "paddle/phi/core/allocator.h"
 #include "paddle/phi/core/dense_tensor.h"
 
 #ifdef INFRT_WITH_GPU
 #include <cuda_runtime.h>
 #endif
 
-namespace paddle {
-namespace platform {
-using DeviceContext = ::phi::DeviceContext;
-}  // namespace platform
-namespace framework {
-using LoDTensor = ::phi::DenseTensor;
-void DeserializeFromStream(std::istream& is,
-                           LoDTensor* tensor,
-                           const platform::DeviceContext& dev_ctx);
-}
-}  // namespace paddle
-
 namespace infrt {
 namespace kernel {
 namespace phi {
@@ -198,6 +189,12 @@ void PrintDenseTensor(::phi::DenseTensor* dense_tensor) {
   auto pb_proto_prog = paddle::LoadProgram(model_path);
   auto main_block = pb_proto_prog->blocks(0);
 
+  ::phi::CPUContext ctx;
+  auto allocator = std::make_unique<backends::CpuPhiAllocator>();
+  const auto* allocator_ptr = allocator.get();
+  ctx.SetAllocator(allocator_ptr);
+  ctx.SetHostAllocator(allocator_ptr);
+  ctx.SetZeroAllocator(allocator_ptr);
   for (auto& var : main_block.vars()) {
     if (var.name() == "feed" || var.name() == "fetch" || !var.persistable())
       continue;
@@ -207,9 +204,7 @@ void PrintDenseTensor(::phi::DenseTensor* dense_tensor) {
       case ::paddle::framework::proto::VarType_Type_LOD_TENSOR: {
         std::unique_ptr<::phi::DenseTensor> tensor{
             std::make_unique<::phi::DenseTensor>()};
-        ::phi::CPUContext ctx;
-        ::paddle::framework::DeserializeFromStream(
-            param_file, tensor.get(), ctx);
+        ::infrt::paddle::DeserializeFromStream(param_file, tensor.get(), ctx);
         map.SetDenseTensor(var.name(), std::move(tensor));
       } break;
       default: {
@@ -249,11 +244,16 @@ void PrintDenseTensor(::phi::DenseTensor* dense_tensor) {
     }
   }
 
+  ::phi::CPUContext ctx;
+  auto allocator = std::make_unique<backends::CpuPhiAllocator>();
+  const auto* allocator_ptr = allocator.get();
+  ctx.SetAllocator(allocator_ptr);
+  ctx.SetHostAllocator(allocator_ptr);
+  ctx.SetZeroAllocator(allocator_ptr);
   for (auto& var : tmp) {
     std::unique_ptr<::phi::DenseTensor> tensor{
         std::make_unique<::phi::DenseTensor>()};
-    ::phi::CPUContext ctx;
-    ::paddle::framework::DeserializeFromStream(param_file, tensor.get(), ctx);
+    ::infrt::paddle::DeserializeFromStream(param_file, tensor.get(), ctx);
     map.SetDenseTensor(var, std::move(tensor));
   }
 
diff --git a/paddle/infrt/paddle/model_parser.cc b/paddle/infrt/paddle/model_parser.cc
index f3de1a630451cc387765040191be8715768be510..da4f8b6420b22dc67d5b1b82d39a86377353b118 100644
--- a/paddle/infrt/paddle/model_parser.cc
+++ b/paddle/infrt/paddle/model_parser.cc
@@ -22,6 +22,10 @@
 #include "paddle/infrt/common/target.h"
 #include "paddle/infrt/common/type.h"
 
+#ifdef INFRT_WITH_PHI
+#include "paddle/phi/common/data_type.h"
+#endif
+
 namespace infrt {
 namespace paddle {
 
@@ -170,5 +174,96 @@ void LoadParam(const std::string &path, _Variable *out, const Target &target) {
   LoadLoDTensor(fin, out, target);
 }
 
+#ifdef INFRT_WITH_PHI
+namespace framework_proto = ::paddle::framework::proto;
+
+inline ::phi::DataType PhiDataType(framework_proto::VarType::Type type) {
+  using Type = framework_proto::VarType::Type;
+  switch (static_cast<int>(type)) {
+    case Type::VarType_Type_BOOL:
+      return ::phi::DataType::BOOL;
+    case Type::VarType_Type_INT8:
+      return ::phi::DataType::INT8;
+    case Type::VarType_Type_UINT8:
+      return ::phi::DataType::UINT8;
+    case Type::VarType_Type_INT16:
+      return ::phi::DataType::INT16;
+    case Type::VarType_Type_INT32:
+      return ::phi::DataType::INT32;
+    case Type::VarType_Type_INT64:
+      return ::phi::DataType::INT64;
+    case Type::VarType_Type_SIZE_T:
+      return ::phi::DataType::UINT64;
+    case Type::VarType_Type_FP16:
+      return ::phi::DataType::FLOAT16;
+    case Type::VarType_Type_FP32:
+      return ::phi::DataType::FLOAT32;
+    case Type::VarType_Type_FP64:
+      return ::phi::DataType::FLOAT64;
+    default:
+      LOG(FATAL) << "unknown data type " << type;
+  }
+  return ::phi::DataType::UNDEFINED;
+}
+
+inline void TensorFromStream(std::istream &is,
+                             ::phi::DenseTensor *tensor,
+                             const ::phi::CPUContext &ctx) {
+  uint32_t version;
+  is.read(reinterpret_cast<char *>(&version), sizeof(version));
+  CHECK_EQ(version, 0U);
+  framework_proto::VarType::TensorDesc desc;
+  {  // int32_t size
+     // proto buffer
+    int32_t size = -1;
+    is.read(reinterpret_cast<char *>(&size), sizeof(size));
+    CHECK_EQ(is.good(), true);
+    CHECK_GE(size, 0);
+    std::unique_ptr<char[]> buf(new char[size]);
+    is.read(reinterpret_cast<char *>(buf.get()), size);
+    CHECK_EQ(desc.ParseFromArray(buf.get(), size), true);
+  }
+  {  // read tensor
+    std::vector<int64_t> dims;
+    dims.reserve(static_cast<size_t>(desc.dims().size()));
+    std::copy(desc.dims().begin(), desc.dims().end(), std::back_inserter(dims));
+    tensor->Resize(::phi::make_ddim(dims));
+    void *buf;
+    size_t size = tensor->numel() * SizeOfType(desc.data_type());
+    ctx.HostAlloc(tensor, PhiDataType(desc.data_type()), size);
+    buf = tensor->data();
+    is.read(static_cast<char *>(buf), size);
+  }
+}
+
+void DeserializeFromStream(std::istream &is,
+                           ::phi::DenseTensor *tensor,
+                           const ::phi::CPUContext &dev_ctx) {
+  {
+    // the 1st field, unit32_t version for LoDTensor
+    uint32_t version;
+    is.read(reinterpret_cast<char *>(&version), sizeof(version));
+    CHECK_EQ(version, 0U);
+  }
+  {
+    // the 2st field, LoD information
+    uint64_t lod_level;
+    is.read(reinterpret_cast<char *>(&lod_level), sizeof(lod_level));
+    auto &lod = *tensor->mutable_lod();
+    lod.resize(lod_level);
+    for (uint64_t i = 0; i < lod_level; ++i) {
+      uint64_t size;
+      is.read(reinterpret_cast<char *>(&size), sizeof(size));
+      std::vector<size_t> tmp(size / sizeof(size_t));
+      is.read(reinterpret_cast<char *>(tmp.data()),
+              static_cast<std::streamsize>(size));
+      lod[i] = tmp;
+    }
+  }
+  // the 3st filed, Tensor
+  TensorFromStream(is, tensor, dev_ctx);
+}
+#endif
+
 }  // namespace paddle
 }  // namespace infrt
diff --git a/paddle/infrt/paddle/model_parser.h b/paddle/infrt/paddle/model_parser.h
index 373f77033dcefa1a81cd8756da859b6d232337a0..5f039ad5d3ad89e25187f1a316e84f8c03df4b63 100644
--- a/paddle/infrt/paddle/model_parser.h
+++ b/paddle/infrt/paddle/model_parser.h
@@ -25,6 +25,11 @@
 #include "paddle/infrt/paddle/scope.h"
 #include "paddle/infrt/paddle/tensor.h"
 
+#ifdef INFRT_WITH_PHI
+#include "paddle/phi/backends/cpu/cpu_context.h"
+#include "paddle/phi/core/dense_tensor.h"
+#endif
+
 namespace infrt {
 namespace paddle {
 namespace framework_proto = ::paddle::framework::proto;
@@ -53,5 +58,11 @@ void TensorFromStream(
     const common::Target& target = common::DefaultHostTarget());
 void ReadBinaryFile(const std::string& filename, std::string* contents);
 
+#ifdef INFRT_WITH_PHI
+void DeserializeFromStream(std::istream& is,
+                           ::phi::DenseTensor* tensor,
+                           const ::phi::CPUContext& dev_ctx);
+#endif
+
 }  // namespace paddle
 }  // namespace infrt