diff --git a/CMakeLists.txt b/CMakeLists.txt
index dc85f683c112e6ec1b5264671c0b6a58a42a31f7..c82a6bf726ff292bb96d2d73ea23f00763dacab3 100755
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -4,8 +4,9 @@ if (ENABLE_OPEN_SRC)
     project(TFAdapter)
 
     set(CMAKE_CXX_STANDARD 11)
-    set(CMAKE_C_FLAGS "-O2 -Wall -fPIC -fstack-protector-all -Wl,-z,relro,-z,now,-z,noexecstack -pipe ${CMAKE_C_FLAGS}")
-    set(CMAKE_CXX_FLAGS "-std=c++11 -O2 -Wall -fPIC -fstack-protector-all -Wl,-z,relro,-z,now,-z,noexecstack -pipe ${CMAKE_CXX_FLAGS}")
+    set(CMAKE_C_FLAGS "-O2 -DNDEBUG -Wall -fPIC -fstack-protector-all -Wl,-z,relro,-z,now,-z,noexecstack -pipe ${CMAKE_C_FLAGS}")
+    set(CMAKE_CXX_FLAGS "-std=c++11 -O2 -DNDEBUG -Wall -fPIC -fstack-protector-all -Wl,-z,relro,-z,now,-z,noexecstack -pipe ${CMAKE_CXX_FLAGS}")
+    add_definitions(-D_GLIBCXX_USE_CXX11_ABI=0)
     # build external prjects
     if(DEFINED ENV{D_PKG_SERVER})
         set(TF_PKG_SERVER $ENV{D_PKG_SERVER})
@@ -18,7 +19,7 @@ if (ENABLE_OPEN_SRC)
     include_directories(${CMAKE_CURRENT_LIST_DIR}/inc)
     include_directories(${CMAKE_CURRENT_LIST_DIR}/inc/external)
     include_directories(${CMAKE_CURRENT_LIST_DIR}/inc/soft_dp)
-
+    
     if (NOT EXISTS ${CMAKE_CURRENT_LIST_DIR}/tools/COMPILE_FLAGS OR NOT EXISTS
             ${CMAKE_CURRENT_LIST_DIR}/tools/LINK_FLAGS OR NOT EXISTS
             ${CMAKE_CURRENT_LIST_DIR}/tools/PYTHON_BIN_PATH OR NOT EXISTS
diff --git a/configure.py b/configure.py
index ce7a3ec12f6b1987ee126df9075632ee4c8aeb72..43f64415f784fb4dfd1991ee155342ae484fd65e 100755
--- a/configure.py
+++ b/configure.py
@@ -111,6 +111,7 @@ def setup_python():
       for flag in compile_args[2:-1]:
         f.write(flag + '\n')
       f.write("-I" + compile_args[-1] + '\n')
+    print('bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb: %s.' % compile_args[1])
     with open(real_config_path('LINK_FLAGS'), 'w') as f:
       f.write(os.path.join(compile_args[1], 'libtensorflow_framework.so.1\n'))
       f.write(os.path.join(compile_args[1], 'python', '_pywrap_tensorflow_internal.so\n'))
@@ -131,15 +132,14 @@ def setup_ascend(env_path):
       break
     elif not os.path.exists(ascend_path):
       print('Invalid ascend path: %s cannot be found.' % ascend_path)
-
+  print('aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa: %s.' % ascend_path)
   with open(real_config_path('LINK_FLAGS'), 'a') as f:
+    f.write(os.path.join(ascend_path, "driver", "lib64", "driver", "libtsdclient.so\n"))
+    f.write(os.path.join(ascend_path, "driver", "lib64", "common", "libc_sec.so\n"))
     f.write(os.path.join(ascend_path, "fwkacllib", "lib64", "libge_runner.so\n"))
     f.write(os.path.join(ascend_path, "fwkacllib", "lib64", "libfmk_parser.so\n"))
     f.write(os.path.join(ascend_path, "fwkacllib", "lib64", "libdatatransfer.so\n"))
     f.write(os.path.join(ascend_path, "fwkacllib", "lib64", "libindextransform.so\n"))
-    f.write(os.path.join(ascend_path, "driver", "lib64", "driver", "libtsdclient.so\n"))
-    f.write(os.path.join(ascend_path, "driver", "lib64", "common", "libc_sec.so\n"))
-
 def setup_swig():
   """Get swig install path."""
   default_swig_path = which('swig')
diff --git a/inc/common/proto/insert_op.proto b/inc/common/proto/insert_op.proto
index e28d65326904bf04df4482372256d2ac5ba5faea..b4c969eb695dafe4cfb0dd4a6e4f3281ca4150ef 100644
--- a/inc/common/proto/insert_op.proto
+++ b/inc/common/proto/insert_op.proto
@@ -38,28 +38,31 @@ message AippOpParams {
 		dynamic = 2;
 	}
 
-	// AIPP模式，区分静态AIPP和动态AIPP
+	// AIPP妯″紡锛屽尯鍒嗛潤鎬丄IPP鍜屽姩鎬丄IPP
 	AippMode aipp_mode = 1;
 
-	// related_input_rank参数为必填，类型为整型，配置范围>=0, <=输入Data算子的个数，默认值为0。
-	// 标识对模型的第几个输入做AIPP处理，例如模型有两个输入，需要对第2个输入做AIPP，则配置related_input_rank为1。
+	// related_input_rank鍙傛暟涓哄繀濉紝绫诲瀷涓烘暣鍨嬶紝閰嶇疆鑼冨洿>=0, <=杈撳叆Data绠楀瓙鐨勪釜鏁帮紝榛樿鍊间负0銆�
+	// 鏍囪瘑瀵规ā鍨嬬殑绗嚑涓緭鍏ュ仛AIPP澶勭悊锛屼緥濡傛ā鍨嬫湁涓や釜杈撳叆锛岄渶瑕佸绗�2涓緭鍏ュ仛AIPP锛屽垯閰嶇疆related_input_rank涓�1銆�
 	uint32 related_input_rank = 2;
 
-	// input_edge_idx参数为可选，类型为整型，配置范围为>=0。
-	// 配置该参数的作用，在于对Data算子不同的输出做不同的AIPP处理，如果该参数没有配置，默认对related_input_rank指定的模型输入的所有输出边做AIPP。
-	// 配置值 <= Data算子输出边的个数。
+    // related_input_name is optional and the top name of data node which inserts aipp
+    string related_input_name = 6;
+
+	// input_edge_idx鍙傛暟涓哄彲閫夛紝绫诲瀷涓烘暣鍨嬶紝閰嶇疆鑼冨洿涓�>=0銆�
+	// 閰嶇疆璇ュ弬鏁扮殑浣滅敤锛屽湪浜庡Data绠楀瓙涓嶅悓鐨勮緭鍑哄仛涓嶅悓鐨凙IPP澶勭悊锛屽鏋滆鍙傛暟娌℃湁閰嶇疆锛岄粯璁ゅrelated_input_rank鎸囧畾鐨勬ā鍨嬭緭鍏ョ殑鎵�鏈夎緭鍑鸿竟鍋欰IPP銆�
+	// 閰嶇疆鍊� <= Data绠楀瓙杈撳嚭杈圭殑涓暟銆�
 	repeated uint32 input_edge_idx = 3;
 
-	// [Begin] 动态AIPP参数，配置静态AIPP时无效
+	// [Begin] 鍔ㄦ�丄IPP鍙傛暟锛岄厤缃潤鎬丄IPP鏃舵棤鏁�
 	uint32 max_src_image_size = 4;
 
-	// 是否支持旋转。默认不支持，开启支持旋转时，会有额外的空间和性能损失
+	// 鏄惁鏀寔鏃嬭浆銆傞粯璁や笉鏀寔锛屽紑鍚敮鎸佹棆杞椂锛屼細鏈夐澶栫殑绌洪棿鍜屾�ц兘鎹熷け
 	bool support_rotation = 5;
 
-	// [End] 动态AIPP参数
+	// [End] 鍔ㄦ�丄IPP鍙傛暟
 
 
-	// [Begin] 静态AIPP参数，配置动态AIPP时无效
+	// [Begin] 闈欐�丄IPP鍙傛暟锛岄厤缃姩鎬丄IPP鏃舵棤鏁�
 	InputFormat input_format = 51;
 	bool csc_switch = 52;
 	float cpadding_value = 53;
@@ -115,7 +118,7 @@ message AippOpParams {
 	repeated int32 input_bias_1 = 43;
 	repeated int32 input_bias_2 = 44;
 
-	// [End] 静态AIPP参数
+	// [End] 闈欐�丄IPP鍙傛暟
 
        // The n number that is used for raw/rgbir data into f16 transformation.
        // The transformation equation is x/(2^n). If set to 0, no transform is performed.
@@ -124,13 +127,13 @@ message AippOpParams {
 
 message MultiShapeOpParams {
 	enum MultiShapeMode {
-		batch      = 0;             //动态batch
-		resolution = 1;             //动态分辨率，扩展用
+		batch      = 0;             //鍔ㄦ�乥atch
+		resolution = 1;             //鍔ㄦ�佸垎杈ㄧ巼锛屾墿灞曠敤
 	}
 
-	MultiShapeMode    mode                      = 1;        //算子模式
-	uint32            related_input_rank        = 2;        //新增算子插入到哪个输入
+	MultiShapeMode    mode                      = 1;        //绠楀瓙妯″紡
+	uint32            related_input_rank        = 2;        //鏂板绠楀瓙鎻掑叆鍒板摢涓緭鍏�
 
 
-	repeated uint32 batch_list   = 11; //batch_list值，batch_list的个数是2到8之间
+	repeated uint32 batch_list   = 11; //batch_list鍊硷紝batch_list鐨勪釜鏁版槸2鍒�8涔嬮棿
 }
diff --git a/inc/common/util/ai_core/common/aicore_util_attr_define.h b/inc/common/util/ai_core/common/aicore_util_attr_define.h
index f3556cb501e6b6decea6e27806046e912e8b83b1..297be0b238e49e4ae1775352677d11bf88fa34ca 100644
--- a/inc/common/util/ai_core/common/aicore_util_attr_define.h
+++ b/inc/common/util/ai_core/common/aicore_util_attr_define.h
@@ -32,6 +32,14 @@ static const std::string L1_OPTIMIZED = "l1_optimized";
 
 static const std::string L2_OPTIMIZED = "l2_optimized";
 
-static const std::string OP_SLICE_INFO = "_op_slice_info";
+static const std::string ATTR_NAME_UNKNOWN_SHAPE = "_unknown_shape";
+
+static const std::string ATTR_NAME_IS_UNKNOWN_GRAPH = "_fe_is_unknown_graph";
+
+static const std::string ATTR_NAME_IS_UNKNOWN_SHAPE_OP = "_fe_is_unknown_shape_op";
+
+static const std::string ATTR_NAME_TVM_CACHE_READ_MODE = "tvm_cache_read_mode";
+
+static const std::string ATTR_NAME_TBE_KERNEL_SIZE = "_tbeKernelSize";
 } // namespace fe
 #endif
diff --git a/inc/common/util/ai_core/common/aicore_util_types.h b/inc/common/util/ai_core/common/aicore_util_types.h
index e57dbca71e33022ab69ce3a97cf12abd8f834e2a..f5e9d229896e66957781c090dd133b1f6a567692 100644
--- a/inc/common/util/ai_core/common/aicore_util_types.h
+++ b/inc/common/util/ai_core/common/aicore_util_types.h
@@ -109,6 +109,10 @@ enum OpImplType {
   EN_RESERVED                        // reserved value
 };
 
+// Dont change the order, only add new mode in the end
+enum L2Mode { EN_L2_CLOSE = 0, EN_L2_BUFFER_OPTIMIZE, EN_L2_CACHE_NORMAL, EN_L2_CACHE_RC };
+enum BufferFusionMode { EN_OPTIMIZE_DISABLE = 0, EN_L2_BUFFER, EN_L2_FUSION};
+
 static const std::map<ge::DataType, uint32_t> DATATYPE_SIZE_MAP{
     {ge::DT_FLOAT, sizeof(float)},
     {ge::DT_FLOAT16, sizeof(int16_t)},
diff --git a/inc/external/ge/ge_api_types.h b/inc/external/ge/ge_api_types.h
index 1c6b7a3ef7dbabcd8a85999ec990671cc5852d8e..d4fccffda72d92c115729d3b0c9d7e6bab93fae6 100644
--- a/inc/external/ge/ge_api_types.h
+++ b/inc/external/ge/ge_api_types.h
@@ -222,6 +222,18 @@ const char *const OPTION_GE_MAX_DUMP_OP_NUM = "ge.maxDumpOpNum";
 // Its value should be "0" or "1", default value is "1"
 const char *const ENABLE_PRINT_OP_PASS = "ge.enablePrintOpPass";
 
+// Configure operator compilation path
+// Its value should be file path, default value is "./"
+const char *const DEBUG_DIR = "ge.debugDir";
+
+// Configure operator compiler cache path
+// Its value should be file path, default value is "./"
+const char *const OP_COMPILER_CACHE_DIR = "ge.op_compiler_cache_dir";
+
+// Configure operator compiler cache mode
+// Its value should be "disable", "enable" or "force", default value is "disable"
+const char *const OP_COMPILER_CACHE_MODE = "ge.op_compiler_cache_mode";
+
 // Configure whether to use single stream.
 // Its value should be "true" or "false", default value is "false"
 const char *const ENABLE_SINGLE_STREAM = "ge.enableSingleStream";
@@ -298,6 +310,9 @@ namespace ir_option {
   static const char *const INPUT_FP16_NODES = ge::INPUT_FP16_NODES.c_str();
   static const char *const LOG_LEVEL = "log";
   static const char *const OPTYPELIST_FOR_IMPLMODE = ge::OPTYPELIST_FOR_IMPLMODE.c_str();
+  static const char *const DEBUG_DIR = ge::DEBUG_DIR;
+  static const char *const OP_COMPILER_CACHE_DIR = ge::OP_COMPILER_CACHE_DIR;
+  static const char *const OP_COMPILER_CACHE_MODE = ge::OP_COMPILER_CACHE_MODE;
 
   // for interface: aclgrphBuildModel
   const std::set<std::string> ir_builder_suppported_options = {
@@ -331,7 +346,10 @@ namespace ir_option {
     FUSION_SWITCH_FILE,
     ENABLE_SMALL_CHANNEL,
     OP_SELECT_IMPL_MODE,
-    OPTYPELIST_FOR_IMPLMODE
+    OPTYPELIST_FOR_IMPLMODE,
+    DEBUG_DIR,
+    OP_COMPILER_CACHE_DIR,
+    OP_COMPILER_CACHE_MODE
   };
 }
 }  // namespace ge
diff --git a/inc/external/graph/gnode.h b/inc/external/graph/gnode.h
index 5465293f8e436f0f645bf6198fe3db9a3bdde96d..882e6d28ead4447856c2470ae81aa81bfc130b2f 100644
--- a/inc/external/graph/gnode.h
+++ b/inc/external/graph/gnode.h
@@ -116,9 +116,9 @@ class GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY GNode {
 
   bool HasAttr(const ge::AscendString &name);
 
-  graphStatus GetSubgraph(uint32_t index, GraphPtr graph) const;
+  graphStatus GetSubgraph(uint32_t index, GraphPtr &graph) const;
 
-  graphStatus GetALLSubgraphs(std::vector<GraphPtr> graph_list) const;
+  graphStatus GetALLSubgraphs(std::vector<GraphPtr> &graph_list) const;
 
  private:
    std::shared_ptr<NodeImpl> impl_;
diff --git a/inc/external/hccl/hccl_types.h b/inc/external/hccl/hccl_types.h
index d67af11e65c24034167ed488eae6af5e701faf68..de982588a3debf67694e08dfa5889fad0cdd4e28 100644
--- a/inc/external/hccl/hccl_types.h
+++ b/inc/external/hccl/hccl_types.h
@@ -19,20 +19,77 @@
 
 #include <stdint.h>
 
-#ifdef _cplusplus
+#ifdef __cplusplus
 extern "C" {
-#endif
+#endif // __cplusplus
 
+/**
+ * @brief HCCL functions return value definition
+ */
 typedef enum {
+    HCCL_SUCCESS = 0,               /**< success */
+    HCCL_E_PARA = 1,                /**< parameter error */
+    HCCL_E_PTR = 2,                 /**< empty pointer */
+    HCCL_E_MEMORY = 3,              /**< memory error */
+    HCCL_E_INTERNAL = 4,            /**< internal error */
+    HCCL_E_NOT_SUPPORT = 5,         /**< not support feature */
+    HCCL_E_NOT_FOUND = 6,           /**< not found specific resource */
+    HCCL_E_UNAVAIL = 7,             /**< resource unavailable */
+    HCCL_E_SYSCALL = 8,             /**< call system interface error */
+    HCCL_E_TIMEOUT = 9,             /**< timeout */
+    HCCL_E_OPEN_FILE_FAILURE = 10,  /**< open file fail */
+    HCCL_E_TCP_CONNECT = 11,        /**< tcp connect fail */
+    HCCL_E_ROCE_CONNECT = 12,       /**< roce connect fail */
+    HCCL_E_TCP_TRANSFER = 13,       /**< tcp transfer fail */
+    HCCL_E_ROCE_TRANSFER = 14,      /**< roce transfer fail */
+    HCCL_E_RUNTIME = 15,            /**< call runtime api fail */
+    HCCL_E_DRV = 16,                /**< call driver api fail */
+    HCCL_E_PROFILING = 17,          /**< call profiling api fail */
+    HCCL_E_CCE = 18,                /**< call cce api fail */
+    HCCL_E_NETWORK = 19,            /**< call network api fail */
+    HCCL_E_RESERVED                 /**< reserved */
 } HcclResult;
 
+/**
+ * @brief handle to HCCL communicator
+ */
+typedef void *HcclComm;
+
+/**
+ * @brief HCCL Reduction opperation
+ */
 typedef enum {
+    HCCL_REDUCE_SUM = 0,    /**< sum */
+    HCCL_REDUCE_PROD = 1,   /**< prod */
+    HCCL_REDUCE_MAX = 2,    /**< max */
+    HCCL_REDUCE_MIN = 3,    /**< min */
+    HCCL_REDUCE_RESERVED    /**< reserved */
 } HcclReduceOp;
 
+/**
+ * @brief HCCL data type
+ */
 typedef enum {
+    HCCL_DATA_TYPE_INT8 = 0,    /**< int8 */
+    HCCL_DATA_TYPE_INT16 = 1,   /**< int16 */
+    HCCL_DATA_TYPE_INT32 = 2,   /**< int32 */
+    HCCL_DATA_TYPE_FP16 = 3,    /**< fp16 */
+    HCCL_DATA_TYPE_FP32 = 4,    /**< fp32 */
+    HCCL_DATA_TYPE_INT64 = 5,    /**< int64 */
+    HCCL_DATA_TYPE_UINT64 = 6,    /**< uint64 */
+    HCCL_DATA_TYPE_RESERVED     /**< reserved */
 } HcclDataType;
 
-#ifdef _cplusplus
+const uint32_t HCCL_ROOT_INFO_BYTES =  4108; // 4108: root info length
+
+/**
+ * @brief HCCL root info
+ */
+typedef struct HcclRootInfoDef {
+    char internal[HCCL_ROOT_INFO_BYTES];
+} HcclRootInfo;
+
+#ifdef __cplusplus
 }
 #endif
 #endif
diff --git a/inc/framework/ge_runtime/task_info.h b/inc/framework/ge_runtime/task_info.h
index e36c43334033b1011cd687b3ae4f7b1df9214f97..2b34c93fa0bfd6da52ff9ca9bccdac47cd9ee073 100644
--- a/inc/framework/ge_runtime/task_info.h
+++ b/inc/framework/ge_runtime/task_info.h
@@ -161,13 +161,12 @@ class TbeTaskInfo : public TaskInfo {
 class AicpuTaskInfo : public TaskInfo {
  public:
   AicpuTaskInfo(const std::string &op_name, uint32_t stream_id, const string &so_name, const std::string &kernel_name,
-                const std::string &node_def, const std::string &ext_info, const std::vector<void *> &input_data_addrs,
+                const std::string &node_def, const std::vector<void *> &input_data_addrs,
                 const std::vector<void *> &output_data_addrs, bool dump_flag)
       : TaskInfo(op_name, stream_id, TaskInfoType::AICPU, dump_flag),
         so_name_(so_name),
         kernel_name_(kernel_name),
         node_def_(node_def),
-        ext_info_(ext_info),
         input_data_addrs_(input_data_addrs),
         output_data_addrs_(output_data_addrs) {}
   ~AicpuTaskInfo() override {}
@@ -177,13 +176,11 @@ class AicpuTaskInfo : public TaskInfo {
   const std::string &node_def() const { return node_def_; }
   const std::vector<void *> &input_data_addrs() const { return input_data_addrs_; }
   const std::vector<void *> &output_data_addrs() const { return output_data_addrs_; }
-  const std::string &ext_info() const { return ext_info_; }
 
  private:
   std::string so_name_;
   std::string kernel_name_;
   std::string node_def_;
-  std::string ext_info_;
   std::vector<void *> input_data_addrs_;
   std::vector<void *> output_data_addrs_;
 };
@@ -308,7 +305,7 @@ class HcclTaskInfo : public TaskInfo {
   int64_t root_id() const { return root_id_; }
   int64_t op_type() const { return op_type_; }
   int64_t data_type() const { return data_type_; }
-  const std::string &group() const { return group_; }
+  const std::string group() const { return group_; }
   std::function<bool(void *, void *)> hcom_bind_model() const { return hcom_bind_model_; }
   std::function<bool(void *)> hcom_unbind_model() const { return hcom_unbind_model_; }
   std::function<bool(std::shared_ptr<HcclTaskInfo>, void *)> hcom_distribute_task() const {
diff --git a/inc/framework/generator/ge_generator.h b/inc/framework/generator/ge_generator.h
index 4902a0213912ee21e075438e257027598ebfafaa..a7bf4a38aafea233decd2b762ac14752f24a2f58 100644
--- a/inc/framework/generator/ge_generator.h
+++ b/inc/framework/generator/ge_generator.h
@@ -88,6 +88,8 @@ class GeGenerator {
                        const string &model_file_name, OpEngineType engine_type,
                        ModelBufferData &model_buff, bool is_offline = true);
 
+  Status CheckForSingleOp(OpDescPtr &op_desc, const vector<GeTensor> &inputs, const vector<GeTensor> &outputs);
+  
   class Impl;
 
   std::shared_ptr<Impl> impl_;
diff --git a/inc/framework/omg/omg_inner_types.h b/inc/framework/omg/omg_inner_types.h
index c48d164944a5e063c6afb7312d07e624f5c4d98b..6cc4afd35b8c693981fcd586ee4e655736a6750c 100644
--- a/inc/framework/omg/omg_inner_types.h
+++ b/inc/framework/omg/omg_inner_types.h
@@ -100,6 +100,8 @@ struct OmgContext {
   std::vector<std::string> net_out_nodes;
   // net out nodes top names(only caffe has top)
   std::vector<std::string> out_top_names;
+  // net data nodes top names(only caffe has top)
+  std::vector<std::string> data_top_names;
   // preferential format used by the entire network
   domiTensorFormat_t net_format = DOMI_TENSOR_RESERVED;
   domi::FrameworkType type = domi::FRAMEWORK_RESERVED;
diff --git a/inc/framework/omg/parser/parser_inner_ctx.h b/inc/framework/omg/parser/parser_inner_ctx.h
index b57420eb1a98464e7417a8dfdb574e38b0fb55d9..31a519ef670736fa84de5685cf23a7673c6478a0 100644
--- a/inc/framework/omg/parser/parser_inner_ctx.h
+++ b/inc/framework/omg/parser/parser_inner_ctx.h
@@ -49,6 +49,8 @@ struct ParserContext {
   std::vector<std::string> user_out_nodes_top_vec;
   // net out nodes (where user_out_nodes or leaf nodes)
   std::vector<std::string> net_out_nodes;
+  // net data nodes top names(only caffe has top)
+  std::vector<std::string> data_top_names;
   // net out nodes top names(only caffe has top)
   std::vector<std::string> out_top_names;
   // Whether to use dynamic batch size or dynamic image size
diff --git a/inc/graph/compute_graph.h b/inc/graph/compute_graph.h
index 90a4c743abdeb147e4e1f4f497590825e9b67f6c..33227bb33fd8750bdb1096147be2bb04bd13b4b1 100644
--- a/inc/graph/compute_graph.h
+++ b/inc/graph/compute_graph.h
@@ -249,11 +249,16 @@ class ComputeGraph : public std::enable_shared_from_this<ComputeGraph>, public A
 
  private:
   graphStatus DFSTopologicalSorting(std::vector<NodePtr> &node_vec, std::map<NodePtr, uint32_t> &map_in_edge_num,
-                                    std::vector<NodePtr> &stack);
+                                    std::vector<NodePtr> &stack, bool reverse);
   graphStatus BFSTopologicalSorting(std::vector<NodePtr> &node_vec, std::map<NodePtr, uint32_t> &map_in_edge_num,
                                     std::deque<NodePtr> &stack);
   graphStatus CollectBreadthOutNode(const NodePtr &node, std::map<NodePtr, uint32_t> &map_in_edge_num,
                                     std::map<string, NodePtr> &breadth_node_map);
+  /// nodes like : (a) <--- (c) ---> (b)
+  /// node a and b have only one parent node c, and a is connected to c firstly
+  /// topo order of DFS is `c, b, a` with `dfs_reverse=false` as default
+  /// in same case, user could get `c, a, b` with `dfs_reverse=true`
+  graphStatus TopologicalSortingGraph(bool dfs_reverse = false);
   graphStatus TopologicalSortingGraph();
   graphStatus SortNodes(std::vector<NodePtr> &stack, std::map<NodePtr, uint32_t> &mapInEdgeNum);
   Vistor<NodePtr> AllGraphNodes(std::vector<std::shared_ptr<ComputeGraph>> &subgraphs) const;
diff --git a/inc/graph/debug/ge_attr_define.h b/inc/graph/debug/ge_attr_define.h
index 334abd1ddf83b061295fa5c8c9b054b76f18b93c..208263822d2d693628af489b3926191db9db2c8d 100644
--- a/inc/graph/debug/ge_attr_define.h
+++ b/inc/graph/debug/ge_attr_define.h
@@ -190,6 +190,7 @@ GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_MOD
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_AUTOMIC_ADD_START;
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_AUTOMIC_ADD_MEM_SIZE;
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_STREAM_LABEL;
+GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_RTS_LABEL_NODE;
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_STREAM_CYCLE_EVENT_FLAG;
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_DYNAMIC_OUTPUT_DIMS;
 
@@ -783,8 +784,6 @@ GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAM
 
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_MODEL_TASK_GEN_VAR_ADDR;
 
-GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_STREAM_LABEL;
-
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_CONTINUOUS_STREAM_LABEL;
 
 GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_MODEL_VAR_SIZE;
diff --git a/inc/graph/ge_local_context.h b/inc/graph/ge_local_context.h
index 36beaa7989dcdf3c3909f4abaed140ee38515091..04060ba48c462a24ab6deae881e93421f90bdd6d 100644
--- a/inc/graph/ge_local_context.h
+++ b/inc/graph/ge_local_context.h
@@ -32,6 +32,12 @@ class GEThreadLocalContext {
   void SetSessionOption(map<std::string, string> options_map);
   void SetGlobalOption(map<std::string, string> options_map);
 
+  map<string, string> GetAllGraphOptions() const;
+  map<string, string> GetAllSessionOptions() const;
+  map<string, string> GetAllGlobalOptions() const;
+  map<string, string> GetAllOptions() const;
+
+
  private:
   map<string, string> graph_options_;
   map<string, string> session_options_;
diff --git a/inc/graph/node.h b/inc/graph/node.h
index 1b33554152cd402135e8f4cfef8593f0dc360f80..66a9ebd102931093039e6e0e0eee33a4a9b0e595 100644
--- a/inc/graph/node.h
+++ b/inc/graph/node.h
@@ -95,7 +95,7 @@ class Node : public std::enable_shared_from_this<Node> {
 
   ComputeGraphPtr GetOwnerComputeGraph() const;
   graphStatus SetOwnerComputeGraph(const ComputeGraphPtr &graph);
-
+  graphStatus SetAnyOwnerComputeGraph(const ComputeGraphPtr &graph);
   Vistor<InDataAnchorPtr> GetAllInDataAnchors() const;
   Vistor<OutDataAnchorPtr> GetAllOutDataAnchors() const;
   uint32_t GetAllInDataAnchorsSize() const;
diff --git a/inc/graph/op_desc.h b/inc/graph/op_desc.h
index 42bf6c67ade7547b7e7f12694dbdf215c758cfcc..8a88ab97cd161c0ab83a79b972865c5e52991751 100644
--- a/inc/graph/op_desc.h
+++ b/inc/graph/op_desc.h
@@ -282,6 +282,8 @@ class OpDesc : public std::enable_shared_from_this<OpDesc>, public AttrHolder {
 
   graphStatus GetSubgraphNameByInstanceName(const std::string &instance_name, std::string &subgraph_name) const;
 
+  graphStatus InferDataSlice();
+
  protected:
   ProtoAttrMapHelper MutableAttrMap() override;
   ConstProtoAttrMapHelper GetAttrMap() const override;
@@ -321,6 +323,7 @@ class OpDesc : public std::enable_shared_from_this<OpDesc>, public AttrHolder {
   std::function<graphStatus(Operator &)> infer_func_ = nullptr;
   std::function<graphStatus(Operator &)> infer_format_func_ = nullptr;
   std::function<graphStatus(Operator &)> verifier_func_ = nullptr;
+  std::function<graphStatus(Operator &)> infer_data_slice_func_ = nullptr;
   string op_kernel_lib_name_;
   string engine_name_;
   friend class OpDescUtils;
diff --git a/inc/graph/utils/graph_utils.h b/inc/graph/utils/graph_utils.h
index 0e7adfe055a51e36a0687249376d7933e92e280b..bdff00b73ea29737f41d0dc6d537222f2e0d35ac 100644
--- a/inc/graph/utils/graph_utils.h
+++ b/inc/graph/utils/graph_utils.h
@@ -145,6 +145,8 @@ class GraphUtils {
 
   static ComputeGraphPtr CreateGraphFromOperator(const string &name, const std::vector<Operator> &inputs);
 
+  static GraphPtr CreateGraphPtrFromComputeGraph(const ComputeGraphPtr compute_graph);
+
   static graphStatus AddEdge(const OutDataAnchorPtr &src, const InDataAnchorPtr &dst);
 
   static graphStatus AddEdge(const OutDataAnchorPtr &src, const Format &src_format, const InDataAnchorPtr &dst,
diff --git a/inc/hccl/base.h b/inc/hccl/base.h
index 94253bf4a72f21ff472ea8cc7ccce18056083b64..8e27d6b809f3206fc3234c79417285bcad93a9a7 100644
--- a/inc/hccl/base.h
+++ b/inc/hccl/base.h
@@ -49,7 +49,7 @@ typedef enum {
     HOROVOD_REDUCE_RESERVED     /**< reserved */
 } HorovodReduceOp;
 
-const u32 HCCL_MAX_SEGMENT_NUM = 8;   // The max number of gradient segments.
+const u32 HCCL_MAX_SEGMENT_NUM = 32;   // The max number of gradient segments.
 
 /**
  * @brief the feature of the model