diff --git a/cmake/anakin_subgraph.cmake b/cmake/anakin_subgraph.cmake
index eb7bce9f3b7a98cb5244ea8d1ce595d5d374cda0..177f34438d6ac8731390e44255072718039bf01c 100644
--- a/cmake/anakin_subgraph.cmake
+++ b/cmake/anakin_subgraph.cmake
@@ -26,6 +26,19 @@ if(ANAKIN_FOUND)
     add_definitions(-DPADDLE_WITH_ANAKIN)
 endif()
 
+if(ANAKIN_FOUND)
+  if (ANAKIN_MLU AND NOT WITH_GPU AND NOT ANAKIN_X86)
+    message(STATUS "Compile with anakin mlu place.")
+    add_definitions(-DANAKIN_MLU_PLACE)
+  elseif(ANAKIN_BM AND NOT WITH_GPU AND NOT ANAKIN_X86)
+    message(STATUS "Compile with anakin bm place.")
+    add_definitions(-DANAKIN_BM_PLACE)
+  elseif(ANAKIN_X86)
+    message(STATUS "Compile with anakin x86 place.")
+    add_definitions(-DANAKIN_X86_PLACE)
+  endif()
+endif()
+
 if(ANAKIN_FOUND AND WITH_GPU AND WITH_DSO)
     message(STATUS "Compile with anakin subgraph.")
     set(ANAKIN_SUBGRAPH ON)
diff --git a/paddle/fluid/inference/api/CMakeLists.txt b/paddle/fluid/inference/api/CMakeLists.txt
index 1921e419383a7b3b657bd4b2ce1b04759d472c65..033224c0f59cbbcd0e8ef2d346ad0aaf4f1d3370 100644
--- a/paddle/fluid/inference/api/CMakeLists.txt
+++ b/paddle/fluid/inference/api/CMakeLists.txt
@@ -68,13 +68,8 @@ cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS analysis_
         ARGS --dirname=${WORD2VEC_MODEL_DIR})
 
 if(ANAKIN_FOUND)
-  if (ANAKIN_MLU AND NOT WITH_GPU AND NOT ANAKIN_X86)
-    message(STATUS "Compile with anakin mlu place.")
-    add_definitions(-DANAKIN_MLU_PLACE)
-  elseif(ANAKIN_X86)
-    message(STATUS "Compile with anakin x86 place.")
-    add_definitions(-DANAKIN_X86_PLACE)
-  endif()
+  # Do not turn warnings into errors.
+  set_source_files_properties(api.cc api_anakin_engine.cc PROPERTIES COMPILE_FLAGS "-Wno-error")
   cc_library(inference_anakin_api SRCS api.cc api_anakin_engine.cc)
   target_link_libraries(inference_anakin_api anakin anakin_saber_common)
   cc_library(inference_anakin_api_shared SHARED SRCS api.cc api_anakin_engine.cc)
diff --git a/paddle/fluid/inference/api/api_anakin_engine.cc b/paddle/fluid/inference/api/api_anakin_engine.cc
index 63d23321ab41eb9f26a4cd79f3ffa799e3795e72..3a7ad2d2db897de2376c8dfe8e6d76ce0bb00a97 100644
--- a/paddle/fluid/inference/api/api_anakin_engine.cc
+++ b/paddle/fluid/inference/api/api_anakin_engine.cc
@@ -34,10 +34,10 @@ extern std::once_flag PaddleInferenceAnakinPredictor<T, P, R>::init_anakin_;
 
 template <typename T, Precision P, OpRunType R>
 void PaddleInferenceAnakinPredictor<T, P, R>::InitEnv() {
-  anakin::TargetWrapper<T>::set_device(this->config_.device_id);
   std::call_once(this->init_anakin_, [this]() {
     anakin::Env<T>::env_init(this->config_.max_stream);
   });
+  anakin::TargetWrapper<T>::set_device(this->config_.device_id);
 }
 template <typename T, Precision P, OpRunType R>
 void PaddleInferenceAnakinPredictor<T, P, R>::InitNet() {
@@ -54,14 +54,19 @@ template <typename T, Precision P, OpRunType R>
 void PaddleInferenceAnakinPredictor<T, P, R>::InitGraph() {
   this->graph_p_ =
       std::make_shared<anakin::graph::Graph<T, anakin::Precision::FP32>>();
-  if (!(this->graph_p_->load(this->config_.model_file))) {
-    LOG(FATAL) << "fail to load graph from " << this->config_.model_file;
+  if (!this->config_.model_file.empty()) {
+    this->graph_p_->load(this->config_.model_file);
+  } else if (this->config_.model_buf_p) {
+    this->graph_p_->load(this->config_.model_buf_p,
+                         this->config_.model_buf_len);
+  } else {
+    LOG(FATAL) << "Model load error.";
   }
   auto inputs = this->graph_p_->get_ins();
   for (auto &input_str : inputs) {
     if (this->config_.init_inputs_shape.find(input_str) ==
         this->config_.init_inputs_shape.end()) {
-      LOG(FATAL) << input_str << " is not implemented.";
+      LOG(FATAL) << input_str << " should be set in init_inputs_shape.";
     }
     std::vector<int> shape =
         this->config_.init_inputs_shape.find(input_str)->second;
@@ -189,6 +194,7 @@ template <typename T, Precision P, OpRunType R>
 bool PaddleInferenceAnakinPredictor<T, P, R>::RunImpl(
     const std::vector<PaddleTensor> &inputs,
     std::vector<PaddleTensor> *output_data) {
+  anakin::TargetWrapper<T>::set_device(this->config_.device_id);
   for (const auto &input : inputs) {
     if (input.dtype != PaddleDType::FLOAT32) {
       LOG(FATAL) << "Only support float type inputs. " << input.name
@@ -321,6 +327,27 @@ void PaddleInferenceAnakinMLUPredictor<P, R>::Predict() {
 }
 #endif
 
+#ifdef ANAKIN_BM_PLACE
+template <Precision P, OpRunType R>
+void PaddleInferenceAnakinBMPredictor<P, R>::OptimizeGraph() {
+  if (!this->graph_p_->fusion_optimize()) {
+    LOG(FATAL) << "Graph optimization error.";
+  }
+}
+template <Precision P, OpRunType R>
+void PaddleInferenceAnakinBMPredictor<P, R>::InitNet() {
+  std::unique_lock<std::mutex> lock(this->mutex_);
+  this->executor_p_ = new anakin::Net<anakin::BM, P, R>();
+  this->executor_p_->fusion_init(*this->graph_p_, this->ctx_p_, true);
+}
+template <Precision P, OpRunType R>
+void PaddleInferenceAnakinBMPredictor<P, R>::Predict() {
+  anakin::TargetWrapper<anakin::BM>::device_sync();
+  this->executor_p_->fusion_prediction();
+  anakin::TargetWrapper<anakin::BM>::device_sync();
+}
+#endif
+
 #ifdef PADDLE_WITH_CUDA
 template class PaddleInferenceAnakinPredictor<
     anakin::NV, anakin::Precision::FP32, ::anakin::OpRunType::ASYNC>;
@@ -333,6 +360,10 @@ template class PaddleInferenceAnakinPredictor<
 template class PaddleInferenceAnakinMLUPredictor<anakin::Precision::FP32,
                                                  ::anakin::OpRunType::SYNC>;
 #endif
+#ifdef ANAKIN_BM_PLACE
+template class PaddleInferenceAnakinBMPredictor<anakin::Precision::FP32,
+                                                ::anakin::OpRunType::ASYNC>;
+#endif
 
 // A factory to help create difference predictor.
 template <>
@@ -361,7 +392,16 @@ CreatePaddlePredictor<contrib::AnakinConfig, PaddleEngineKind::kAnakin>(
             config));
   }
 #endif
-  LOG(FATAL) << "Anakin Predictor create on unknown platform.";
+#ifdef ANAKIN_BM_PLACE
+  if (config.target_type == contrib::AnakinConfig::BM) {
+    return std::unique_ptr<PaddlePredictor>(
+        new PaddleInferenceAnakinBMPredictor<anakin::Precision::FP32,
+                                             ::anakin::OpRunType::ASYNC>(
+            config));
+  }
+#endif
+  LOG(FATAL) << "Anakin Predictor create on unknown platform: "
+             << config.target_type;
   return nullptr;
 }
 template <typename T, Precision P, OpRunType R>
diff --git a/paddle/fluid/inference/api/api_anakin_engine.h b/paddle/fluid/inference/api/api_anakin_engine.h
index 0f0d7febe2ed7331830227c95fed31ac496fa129..17a4c9fa4f6a961db1bcc0a35cc12613f7a8be48 100644
--- a/paddle/fluid/inference/api/api_anakin_engine.h
+++ b/paddle/fluid/inference/api/api_anakin_engine.h
@@ -92,4 +92,19 @@ class PaddleInferenceAnakinMLUPredictor final
   void Predict() override;
 };
 #endif
+
+#ifdef ANAKIN_BM_PLACE
+template <Precision P, OpRunType R>
+class PaddleInferenceAnakinBMPredictor final
+    : public PaddleInferenceAnakinPredictor<anakin::BM, P, R> {
+ public:
+  explicit PaddleInferenceAnakinBMPredictor(const AnakinConfig& config) {
+    this->ResetConfig(config);
+    this->InitPredictor();
+  }
+  void OptimizeGraph() override;
+  void InitNet() override;
+  void Predict() override;
+};
+#endif
 }  // namespace paddle
diff --git a/paddle/fluid/inference/api/paddle_anakin_config.h b/paddle/fluid/inference/api/paddle_anakin_config.h
index 7c0e2f06ff417d22b2adfec3387e9bcbc7e5e81e..e9af13f526a2341a6e02c6979493e995c72dfa98 100644
--- a/paddle/fluid/inference/api/paddle_anakin_config.h
+++ b/paddle/fluid/inference/api/paddle_anakin_config.h
@@ -25,7 +25,7 @@ namespace paddle {
 namespace contrib {
 // Configurations for Anakin engine.
 struct AnakinConfig : public PaddlePredictor::Config {
-  enum TargetType { NVGPU = 0, X86, MLU };
+  enum TargetType { NVGPU = 0, X86, MLU, BM };
   int device_id{0};
   std::string model_file;
   std::map<std::string, std::vector<int>> init_inputs_shape;
@@ -34,6 +34,8 @@ struct AnakinConfig : public PaddlePredictor::Config {
   int max_stream{4};
   int data_stream_id{0};
   int compute_stream_id{0};
+  char* model_buf_p{nullptr};
+  size_t model_buf_len{0};
   TargetType target_type;
 #ifdef ANAKIN_MLU_PLACE
   int model_parallel{8};