diff --git a/CMakeLists.txt b/CMakeLists.txt
index 9ad69738eb2ac21d6ff2624f11d17a38410d5c1f..26d94384a9150735aa8341fd8a18cb039895ff91 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -71,7 +71,8 @@ option(ANAKIN_BUILD_CROSS_PLANTFORM "Build anakin lib for any nvidia device plan
 option(WITH_GRPC     "Use grpc as the default rpc framework"            ${WITH_DISTRIBUTE})
 option(WITH_BRPC_RDMA     "Use brpc rdma as the rpc protocal"           OFF)
 option(ON_INFER         "Turn on inference optimization."               OFF)
-option(WITH_INFERENCE_API_TEST   "Test fluid inference high-level api interface"  OFF)
+option(WITH_INFERENCE_API_TEST   "Test fluid inference C++ high-level api interface"  OFF)
+option(WITH_HIGH_LEVEL_API_TEST   "Test fluid python high-level api interface"  OFF)
 option(WITH_SYSTEM_BLAS   "Use system blas library"           OFF)
 option(PY_VERSION       "Compile PaddlePaddle with python3 support"     ${PY_VERSION})
 option(WITH_FAST_MATH   "Make use of fast math library, might affect the precision to some extent" ON)
diff --git a/cmake/external/protobuf.cmake b/cmake/external/protobuf.cmake
index 69da9b98198de358348621ecdb444f2f81c7757f..09eb437aede4364f8aa285d5296f21cd8460fca1 100644
--- a/cmake/external/protobuf.cmake
+++ b/cmake/external/protobuf.cmake
@@ -221,6 +221,7 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
             -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
             -DCMAKE_INSTALL_PREFIX=${PROTOBUF_INSTALL_DIR}
             -DCMAKE_INSTALL_LIBDIR=lib
+            -DBUILD_SHARED_LIBS=OFF
         CMAKE_CACHE_ARGS
             -DCMAKE_INSTALL_PREFIX:PATH=${PROTOBUF_INSTALL_DIR}
             -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec
index c0df4c2dc6cd46a90d4205b26123fb6c44fcaf88..bf39325cc9bfb258051ec1a7fc7f5eb139c60133 100644
--- a/paddle/fluid/API.spec
+++ b/paddle/fluid/API.spec
@@ -118,6 +118,8 @@ paddle.fluid.layers.reduce_mean (ArgSpec(args=['input', 'dim', 'keep_dim', 'name
 paddle.fluid.layers.reduce_max (ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)), ('document', '66a622db727551761ce4eb73eaa7f6a4'))
 paddle.fluid.layers.reduce_min (ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)), ('document', 'd50ac552b5d131468ed466d08bb2d38c'))
 paddle.fluid.layers.reduce_prod (ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)), ('document', 'fcd8301a0ce15f219c7a4bcd0c1e8eca'))
+paddle.fluid.layers.reduce_all (ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)), ('document', '646ca4d4a2cc16084f59de44b6927eca'))
+paddle.fluid.layers.reduce_any (ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)), ('document', 'f36661060aeeaf6c6b1331e41b3726fa'))
 paddle.fluid.layers.sequence_first_step (ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None), ('document', '2b290d3d77882bfe9bb8d331cac8cdd3'))
 paddle.fluid.layers.sequence_last_step (ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None), ('document', 'c16a892f44f7fe71bfa5afc32d3f34ce'))
 paddle.fluid.layers.sequence_slice (ArgSpec(args=['input', 'offset', 'length', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'fdcea0e8b5bc7d8d4b1b072c521014e6'))
@@ -125,7 +127,7 @@ paddle.fluid.layers.dropout (ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed
 paddle.fluid.layers.split (ArgSpec(args=['input', 'num_or_sections', 'dim', 'name'], varargs=None, keywords=None, defaults=(-1, None)), ('document', '652625345c2acb900029c78cc75f8aa6'))
 paddle.fluid.layers.ctc_greedy_decoder (ArgSpec(args=['input', 'blank', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'ebbf2adbd79683dc93db03454dfa18c2'))
 paddle.fluid.layers.edit_distance (ArgSpec(args=['input', 'label', 'normalized', 'ignored_tokens'], varargs=None, keywords=None, defaults=(True, None)), ('document', '97f0262f97602644c83142789d784571'))
-paddle.fluid.layers.l2_normalize (ArgSpec(args=['x', 'axis', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(1e-12, None)), ('document', '6e428384ce6a77207fa2c70d9f011990'))
+paddle.fluid.layers.l2_normalize (ArgSpec(args=['x', 'axis', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(1e-12, None)), ('document', '35c6a241bcc1a1fc89508860d82ad62b'))
 paddle.fluid.layers.matmul (ArgSpec(args=['x', 'y', 'transpose_x', 'transpose_y', 'alpha', 'name'], varargs=None, keywords=None, defaults=(False, False, 1.0, None)), ('document', 'b4cbe1ac451005df6dad12e9ffdccca9'))
 paddle.fluid.layers.topk (ArgSpec(args=['input', 'k', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'd3570c02f71bcd78e60b3f31dc8f5b32'))
 paddle.fluid.layers.warpctc (ArgSpec(args=['input', 'label', 'blank', 'norm_by_times', 'use_cudnn'], varargs=None, keywords=None, defaults=(0, False, False)), ('document', 'aaba49c038ba927f0a8e45c0c9a686ab'))
@@ -204,6 +206,7 @@ paddle.fluid.layers.gaussian_random_batch_size_like (ArgSpec(args=['input', 'sha
 paddle.fluid.layers.sum (ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None), ('document', 'a418e3ccb5e2ac21bd60f5cc221d5860'))
 paddle.fluid.layers.slice (ArgSpec(args=['input', 'axes', 'starts', 'ends'], varargs=None, keywords=None, defaults=None), ('document', '01dbb91e7c74cb11336cd531013de51a'))
 paddle.fluid.layers.shape (ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None), ('document', '17db0f814eb7bb5a3fac1ca6e60e16d8'))
+paddle.fluid.layers.rank (ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None), ('document', 'ee1386c42ecc8f424fe3fb21862fefc2'))
 paddle.fluid.layers.logical_and (ArgSpec(args=['x', 'y', 'out', 'name'], varargs=None, keywords=None, defaults=(None, None)), ('document', 'cdcf20c494c92060d10feb9374532f42'))
 paddle.fluid.layers.logical_or (ArgSpec(args=['x', 'y', 'out', 'name'], varargs=None, keywords=None, defaults=(None, None)), ('document', '0eae3f726a4afe590757552fa3ced012'))
 paddle.fluid.layers.logical_xor (ArgSpec(args=['x', 'y', 'out', 'name'], varargs=None, keywords=None, defaults=(None, None)), ('document', 'b0daaa3fa4a0aa62f9b58c43d959eb25'))
@@ -236,7 +239,7 @@ paddle.fluid.layers.huber_loss (ArgSpec(args=['input', 'label', 'delta'], vararg
 paddle.fluid.layers.kldiv_loss (ArgSpec(args=['x', 'target', 'reduction', 'name'], varargs=None, keywords=None, defaults=('mean', None)), ('document', '776d536cac47c89073abc7ee524d5aec'))
 paddle.fluid.layers.tree_conv (ArgSpec(args=['nodes_vector', 'edge_set', 'output_size', 'num_filters', 'max_depth', 'act', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(1, 2, 'tanh', None, None, None)), ('document', '34ea12ac9f10a65dccbc50100d12e607'))
 paddle.fluid.layers.npair_loss (ArgSpec(args=['anchor', 'positive', 'labels', 'l2_reg'], varargs=None, keywords=None, defaults=(0.002,)), ('document', '46994d10276dd4cb803b4062b5d14329'))
-paddle.fluid.layers.pixel_shuffle (ArgSpec(args=['x', 'upscale_factor'], varargs=None, keywords=None, defaults=None), ('document', 'ad669cdf83e72a69ebc5ed79e36486de'))
+paddle.fluid.layers.pixel_shuffle (ArgSpec(args=['x', 'upscale_factor'], varargs=None, keywords=None, defaults=None), ('document', '731b21c62a4add60a33bd76d802ffc5c'))
 paddle.fluid.layers.fsp_matrix (ArgSpec(args=['x', 'y'], varargs=None, keywords=None, defaults=None), ('document', 'b76ccca3735bea4a58a0dbf0d77c5393'))
 paddle.fluid.layers.data (ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True)), ('document', '33bbd42027d872b3818b3d64ec52e139'))
 paddle.fluid.layers.open_files (ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None)), ('document', 'b1ae2e1cc0750e58726374061ea90ecc'))
@@ -272,6 +275,7 @@ paddle.fluid.layers.has_inf (ArgSpec(args=['x'], varargs=None, keywords=None, de
 paddle.fluid.layers.has_nan (ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None), ('document', '2e53e83127dbfd86e7098bdfe9a549e8'))
 paddle.fluid.layers.isfinite (ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None), ('document', '0a437011c3906079fd8947ed3e52d292'))
 paddle.fluid.layers.range (ArgSpec(args=['start', 'end', 'step', 'dtype'], varargs=None, keywords=None, defaults=None), ('document', '2ec937ede953ded2fdff2675883900bb'))
+paddle.fluid.layers.linspace (ArgSpec(args=['start', 'stop', 'num', 'dtype'], varargs=None, keywords=None, defaults=None), ('document', '495e21e9a848c2d075a102802fc67756'))
 paddle.fluid.layers.While.__init__ (ArgSpec(args=['self', 'cond', 'is_test', 'name'], varargs=None, keywords=None, defaults=(False, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.layers.While.block (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.layers.Switch.__init__ (ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
@@ -297,12 +301,12 @@ paddle.fluid.layers.DynamicRNN.static_input (ArgSpec(args=['self', 'x'], varargs
 paddle.fluid.layers.DynamicRNN.step_input (ArgSpec(args=['self', 'x', 'level'], varargs=None, keywords=None, defaults=(0,)), ('document', '7568c5ac7622a10288d3307a94134655'))
 paddle.fluid.layers.DynamicRNN.update_memory (ArgSpec(args=['self', 'ex_mem', 'new_mem'], varargs=None, keywords=None, defaults=None), ('document', '5d83987da13b98363d6a807a52d8024f'))
 paddle.fluid.layers.StaticRNN.__init__ (ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.layers.StaticRNN.memory (ArgSpec(args=['self', 'init', 'shape', 'batch_ref', 'init_value', 'init_batch_dim_idx', 'ref_batch_dim_idx'], varargs=None, keywords=None, defaults=(None, None, None, 0.0, 0, 1)), ('document', 'c24e368e23afac1ed91a78a639d7a9c7'))
-paddle.fluid.layers.StaticRNN.output (ArgSpec(args=['self'], varargs='outputs', keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.layers.StaticRNN.step (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.layers.StaticRNN.step_input (ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.layers.StaticRNN.step_output (ArgSpec(args=['self', 'o'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.layers.StaticRNN.update_memory (ArgSpec(args=['self', 'mem', 'var'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
+paddle.fluid.layers.StaticRNN.memory (ArgSpec(args=['self', 'init', 'shape', 'batch_ref', 'init_value', 'init_batch_dim_idx', 'ref_batch_dim_idx'], varargs=None, keywords=None, defaults=(None, None, None, 0.0, 0, 1)), ('document', '72530f299d6451a567cf4a12dc3fb1ff'))
+paddle.fluid.layers.StaticRNN.output (ArgSpec(args=['self'], varargs='outputs', keywords=None, defaults=None), ('document', 'df6ceab6e6c9bd31e97914d7e7538137'))
+paddle.fluid.layers.StaticRNN.step (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6d3e0a5d9aa519a9773a36e1620ea9b7'))
+paddle.fluid.layers.StaticRNN.step_input (ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None), ('document', '903387ec11f3d0bf46821d31a68cffa5'))
+paddle.fluid.layers.StaticRNN.step_output (ArgSpec(args=['self', 'o'], varargs=None, keywords=None, defaults=None), ('document', '252890d4c3199a7623ab8667e13fd837'))
+paddle.fluid.layers.StaticRNN.update_memory (ArgSpec(args=['self', 'mem', 'var'], varargs=None, keywords=None, defaults=None), ('document', '7a0000520f179f35239956a5ba55119f'))
 paddle.fluid.layers.reorder_lod_tensor_by_rank (ArgSpec(args=['x', 'rank_table'], varargs=None, keywords=None, defaults=None), ('document', '3545f529ef04e8f6ecb76b47fa3df01a'))
 paddle.fluid.layers.Print (ArgSpec(args=['input', 'first_n', 'message', 'summarize', 'print_tensor_name', 'print_tensor_type', 'print_tensor_shape', 'print_tensor_lod', 'print_phase'], varargs=None, keywords=None, defaults=(-1, None, -1, True, True, True, True, 'both')), ('document', '5fef91b0e21c93610785f2b1f7161732'))
 paddle.fluid.layers.is_empty (ArgSpec(args=['x', 'cond'], varargs=None, keywords=None, defaults=(None,)), ('document', 'bbe578dbb49ad13e15b014e98c22b519'))
@@ -361,8 +365,7 @@ paddle.fluid.layers.inverse_time_decay (ArgSpec(args=['learning_rate', 'decay_st
 paddle.fluid.layers.polynomial_decay (ArgSpec(args=['learning_rate', 'decay_steps', 'end_learning_rate', 'power', 'cycle'], varargs=None, keywords=None, defaults=(0.0001, 1.0, False)), ('document', '882634f420f626642f0874481263da40'))
 paddle.fluid.layers.piecewise_decay (ArgSpec(args=['boundaries', 'values'], varargs=None, keywords=None, defaults=None), ('document', 'c717d9d1d78a53c809d01b8bc56f3cae'))
 paddle.fluid.layers.noam_decay (ArgSpec(args=['d_model', 'warmup_steps'], varargs=None, keywords=None, defaults=None), ('document', 'd9a95746353fd574be36dc28d8726c28'))
-paddle.fluid.layers.append_LARS (ArgSpec(args=['params_grads', 'learning_rate', 'weight_decay'], varargs=None, keywords=None, defaults=None), ('document', 'd24fa1e7d62ac8a534fc6a86002f84f8'))
-paddle.fluid.layers.cosine_decay (ArgSpec(args=['learning_rate', 'step_each_epoch', 'epochs'], varargs=None, keywords=None, defaults=None), ('document', '9588c64c26ffaef3c466e404a6af9d9b'))
+paddle.fluid.layers.cosine_decay (ArgSpec(args=['learning_rate', 'step_each_epoch', 'epochs'], varargs=None, keywords=None, defaults=None), ('document', 'f8b2727bccf0f368c997d7cf05847e49'))
 paddle.fluid.layers.linear_lr_warmup (ArgSpec(args=['learning_rate', 'warmup_steps', 'start_lr', 'end_lr'], varargs=None, keywords=None, defaults=None), ('document', '2ef3f5ca5cd71ea4217c418e5a7a0565'))
 paddle.fluid.contrib.InitState.__init__ (ArgSpec(args=['self', 'init', 'shape', 'value', 'init_boot', 'need_reorder', 'dtype'], varargs=None, keywords=None, defaults=(None, None, 0.0, None, False, 'float32')), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.contrib.StateCell.__init__ (ArgSpec(args=['self', 'inputs', 'states', 'out_state', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
diff --git a/paddle/fluid/framework/data_feed.cc b/paddle/fluid/framework/data_feed.cc
index b5f7e6c22405d6928f0e423458d6cd720f2d09a8..365c80da34eb287f50d2f0dcbf3844001ab43ec8 100644
--- a/paddle/fluid/framework/data_feed.cc
+++ b/paddle/fluid/framework/data_feed.cc
@@ -72,7 +72,6 @@ bool DataFeed::PickOneFile(std::string* filename) {
   }
   VLOG(3) << "file_idx_=" << *file_idx_;
   *filename = filelist_[(*file_idx_)++];
-  // LOG(ERROR) << "pick file:" << *filename;
   return true;
 }
 
@@ -466,6 +465,17 @@ void MultiSlotDataFeed::Init(
     if (slot.is_used()) {
       use_slots_.push_back(all_slots_[i]);
       use_slots_is_dense_.push_back(slot.is_dense());
+      std::vector<int> local_shape;
+      if (slot.is_dense()) {
+        // for batch size holder if is_dense
+        if (slot.shape(0) > 0) {
+          local_shape.push_back(0);
+        }
+      }
+      for (size_t i = 0; i < slot.shape_size(); ++i) {
+        local_shape.push_back(slot.shape(i));
+      }
+      use_slots_shape_.push_back(local_shape);
     }
   }
   feed_vec_.resize(use_slots_.size());
@@ -752,8 +762,8 @@ void MultiSlotDataFeed::PutToFeedVec(
     LoD data_lod{offset};
     feed_vec_[i]->set_lod(data_lod);
     if (use_slots_is_dense_[i]) {
-      int dim = total_instance / batch_size_;
-      feed_vec_[i]->Resize({batch_size_, dim});
+      use_slots_shape_[i][0] = batch_size_;
+      feed_vec_[i]->Resize(framework::make_ddim(use_slots_shape_[i]));
     }
   }
 #endif
@@ -785,6 +795,16 @@ void MultiSlotInMemoryDataFeed::Init(
     if (slot.is_used()) {
       use_slots_.push_back(all_slots_[i]);
       use_slots_is_dense_.push_back(slot.is_dense());
+      std::vector<int> local_shape;
+      if (slot.is_dense()) {
+        if (slot.shape(0) > 0) {
+          local_shape.push_back(0);
+        }
+      }
+      for (size_t i = 0; i < slot.shape_size(); ++i) {
+        local_shape.push_back(slot.shape(i));
+      }
+      use_slots_shape_.push_back(local_shape);
     }
   }
   feed_vec_.resize(use_slots_.size());
@@ -940,8 +960,8 @@ void MultiSlotInMemoryDataFeed::PutToFeedVec(
     LoD data_lod{offset};
     feed_vec_[i]->set_lod(data_lod);
     if (use_slots_is_dense_[i]) {
-      int dim = total_instance / batch_size_;
-      feed_vec_[i]->Resize({batch_size_, dim});
+      use_slots_shape_[i][0] = batch_size_;
+      feed_vec_[i]->Resize(framework::make_ddim(use_slots_shape_[i]));
     }
   }
 #endif
diff --git a/paddle/fluid/framework/data_feed.h b/paddle/fluid/framework/data_feed.h
index 648c874a0b8763b18118e18adf3b3e93acfd104b..d098c7858a98c644bd3cad78d3cf1e3b35ca026b 100644
--- a/paddle/fluid/framework/data_feed.h
+++ b/paddle/fluid/framework/data_feed.h
@@ -142,6 +142,7 @@ class DataFeed {
   // object)
   std::vector<std::string> all_slots_;
   std::vector<std::string> all_slots_type_;
+  std::vector<std::vector<int>> use_slots_shape_;
   std::vector<int>
       use_slots_index_;  // -1: not used; >=0: the index of use_slots_
 
diff --git a/paddle/fluid/framework/data_feed.proto b/paddle/fluid/framework/data_feed.proto
index 77911306299b77748a2ad9437d49680748885003..03996e0e20a1729ee300a5ad37abc325876930b7 100644
--- a/paddle/fluid/framework/data_feed.proto
+++ b/paddle/fluid/framework/data_feed.proto
@@ -19,6 +19,7 @@ message Slot {
   required string type = 2;
   optional bool is_dense = 3 [ default = false ];
   optional bool is_used = 4 [ default = false ];
+  repeated int32 shape = 5; // we can define N-D Tensor
 }
 
 message MultiSlotDesc { repeated Slot slots = 1; }
diff --git a/paddle/fluid/framework/details/async_ssa_graph_executor.cc b/paddle/fluid/framework/details/async_ssa_graph_executor.cc
index e9aad5d264d1745662848d1ba313b573d0974cb7..7f63c07b18f7c6147670656dfc567f8f2ae8429a 100644
--- a/paddle/fluid/framework/details/async_ssa_graph_executor.cc
+++ b/paddle/fluid/framework/details/async_ssa_graph_executor.cc
@@ -64,9 +64,12 @@ void ProcessGraph(std::vector<ir::Graph *> graphs, Scope *scope) {
               node->Op()->GetNullableAttr("epmap"));
           auto height_section = boost::get<std::vector<int64_t>>(
               node->Op()->GetNullableAttr("sections"));
+          auto trainer_id =
+              boost::get<int>(node->Op()->GetNullableAttr("trainer_id"));
           send_varname_to_ctx[send_var_name] =
               operators::distributed::RpcContext(send_var_name, send_varnames,
-                                                 epmap, height_section);
+                                                 epmap, height_section,
+                                                 trainer_id);
           VLOG(3) << "find and init an send op: "
                   << send_varname_to_ctx[send_var_name];
         } else if (node->Name() == "recv") {
@@ -75,9 +78,11 @@ void ProcessGraph(std::vector<ir::Graph *> graphs, Scope *scope) {
               node->Op()->GetNullableAttr("recv_varnames"));
           auto epmap = boost::get<std::vector<std::string>>(
               node->Op()->GetNullableAttr("epmap"));
+          auto trainer_id =
+              boost::get<int>(node->Op()->GetNullableAttr("trainer_id"));
           recv_varname_to_ctx[recv_var_name] =
               operators::distributed::RpcContext(recv_var_name, recv_varnames,
-                                                 epmap, {});
+                                                 epmap, {}, trainer_id);
           nodes_to_delete.push_back(node);
           VLOG(3) << "find and remove an recv op: "
                   << recv_varname_to_ctx[recv_var_name];
diff --git a/paddle/fluid/framework/details/build_strategy.cc b/paddle/fluid/framework/details/build_strategy.cc
index afe5078bf80d00b595789a5f45d91a5e7a8dfce6..196603bbff1db79e46ebbe8b18f1092fcbaac7f9 100644
--- a/paddle/fluid/framework/details/build_strategy.cc
+++ b/paddle/fluid/framework/details/build_strategy.cc
@@ -101,8 +101,6 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
                "mode.";
         strategy_.fuse_all_optimizer_ops_ = false;
       } else {
-        VLOG(10) << "Add alloc_continuous_space_for_grad_pass";
-        AppendPass("alloc_continuous_space_for_grad_pass");
         // NOTE: fuse_all_xx_ops will count the number of xx operator first,
         // if the number is zero, fuse_all_reduce_ops will do nothing.
         // Currently, only one type of optimization algorithm can be fused.
@@ -150,6 +148,11 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
       AppendPass("runtime_context_cache_pass");
     }
 
+    if (strategy_.cache_expected_kernel_) {
+      VLOG(10) << "Add expected_kernel_cache_pass";
+      AppendPass("expected_kernel_cache_pass");
+    }
+
     AppendMultiDevPass(strategy_);
 
     if (strategy_.fuse_all_reduce_ops_) {
@@ -337,3 +340,4 @@ USE_PASS(fuse_adam_op_pass);
 USE_PASS(fuse_sgd_op_pass);
 USE_PASS(fuse_all_reduce_op_pass);
 USE_PASS(runtime_context_cache_pass);
+USE_PASS(expected_kernel_cache_pass);
diff --git a/paddle/fluid/framework/details/build_strategy.h b/paddle/fluid/framework/details/build_strategy.h
index 8aa444a30c0f7f1f5c19d54cf248f86c3e3b3cf3..b1601cfbcd5e9c66f1bbecd1f6fe10bc279cea26 100644
--- a/paddle/fluid/framework/details/build_strategy.h
+++ b/paddle/fluid/framework/details/build_strategy.h
@@ -83,11 +83,11 @@ struct BuildStrategy {
 
   bool sync_batch_norm_{false};
 
-  bool memory_optimize_{true};
-  // TODO(dzhwinter):
-  // make enable_inplace, memory_optimize_
-  // memory_early_delete_ true by default
-  bool enable_inplace_{true};
+  // FIXME(liuwei1031) disable memory_optimzie and enable_inplace in 1.4
+  // to open them by default, we need to solve the fetch variable issue
+  bool memory_optimize_{false};
+
+  bool enable_inplace_{false};
 
   bool enable_sequential_execution_{false};
 
@@ -108,6 +108,7 @@ struct BuildStrategy {
   bool remove_unnecessary_lock_{true};
 
   bool cache_runtime_context_{false};
+  bool cache_expected_kernel_{true};
 
   // NOTE:
   // Before you add new options, think if it's a general strategy that works
diff --git a/paddle/fluid/framework/details/fuse_adam_op_pass.cc b/paddle/fluid/framework/details/fuse_adam_op_pass.cc
index 0ef75e319244e2ccc63dfa3f93f0cd764cf67633..f95d93fd5575ae538274c4c0322cf661c631849a 100644
--- a/paddle/fluid/framework/details/fuse_adam_op_pass.cc
+++ b/paddle/fluid/framework/details/fuse_adam_op_pass.cc
@@ -24,7 +24,7 @@ namespace details {
 const std::string FuseAdamOpPass::GetOpType() const { return "adam"; }
 
 const std::vector<std::string> FuseAdamOpPass::GetAuxiliaryVarNames() const {
-  return {"Param", "Moment1", "Moment2", "Beta1Pow", "Beta2Pow"};
+  return {"Moment1", "Moment2", "Beta1Pow", "Beta2Pow"};
 }
 
 void FuseAdamOpPass::FuseOptimizerOps(
@@ -77,16 +77,16 @@ void FuseAdamOpPass::FuseAdamOps(
   VLOG(10) << "Insert adam to graph ";
   OpDesc adam_desc(adam_ops[0]->Op()->Block());
   adam_desc.SetType("adam");
-  adam_desc.SetInput("Param", {fused_vars_name.at("Param")});
-  adam_desc.SetInput("Grad", {fused_vars_name.at("Grad")});
+  adam_desc.SetInput(kParam, {fused_vars_name.at(kParam)});
+  adam_desc.SetInput(kGrad, {fused_vars_name.at(kGrad)});
   adam_desc.SetInput("Moment1", {fused_vars_name.at("Moment1")});
   adam_desc.SetInput("Moment2", {fused_vars_name.at("Moment2")});
   // TODO(zcd): The LearningRate, Beta1Pow, Beta2Pow should be equal.
-  adam_desc.SetInput("LearningRate", adam_ops[0]->Op()->Input("LearningRate"));
+  adam_desc.SetInput(kLearningRate, adam_ops[0]->Op()->Input(kLearningRate));
   adam_desc.SetInput("Beta1Pow", adam_ops[0]->Op()->Input("Beta1Pow"));
   adam_desc.SetInput("Beta2Pow", adam_ops[0]->Op()->Input("Beta2Pow"));
 
-  adam_desc.SetOutput("ParamOut", {fused_vars_name.at("Param")});
+  adam_desc.SetOutput("ParamOut", {fused_vars_name.at(kParam)});
   adam_desc.SetOutput("Moment1Out", {fused_vars_name.at("Moment1")});
   adam_desc.SetOutput("Moment2Out", {fused_vars_name.at("Moment2")});
   adam_desc.SetAttr("beta1", beta1);
diff --git a/paddle/fluid/framework/details/fuse_optimizer_op_pass.cc b/paddle/fluid/framework/details/fuse_optimizer_op_pass.cc
index b49f095d428a017dd1a3bed2788a048af9afa6bb..25aa3019d102293725d836cf1f8e9fce8462408b 100644
--- a/paddle/fluid/framework/details/fuse_optimizer_op_pass.cc
+++ b/paddle/fluid/framework/details/fuse_optimizer_op_pass.cc
@@ -29,7 +29,9 @@ void FuseOptimizerOpPass::ApplyImpl(ir::Graph *graph) const {
   auto &local_scopes = Get<const std::vector<Scope *>>(kLocalScopes);
 
   const std::string fuse_op_type = GetOpType();
-  const std::vector<std::string> aux_var_names = GetAuxiliaryVarNames();
+  std::vector<std::string> aux_var_names = GetAuxiliaryVarNames();
+  aux_var_names.emplace_back(kParam);
+  aux_var_names.emplace_back(kGrad);
 
   // Step 1: Get the specified op and auxiliary variables.
   std::vector<ir::Node *> topo_nodes = ir::TopologySortOperations(result);
@@ -61,7 +63,7 @@ void FuseOptimizerOpPass::ApplyImpl(ir::Graph *graph) const {
     result.Set(kFusedVars, new FusedVars);
   }
   std::unordered_map<std::string, std::string> fused_vars_name;
-  fused_vars_name.reserve(aux_var_names.size() + 1);
+  fused_vars_name.reserve(aux_var_names.size());
   auto &fused_var_set = result.Get<FusedVars>(kFusedVars);
   const std::string prefix(kFusedVarNamePrefix);
   // NOTE: the fused_var_name should be unique.
@@ -75,39 +77,103 @@ void FuseOptimizerOpPass::ApplyImpl(ir::Graph *graph) const {
   }
 
   // Step 3: Get the fused Gradient's name
-  auto &params_grads = result.Get<ParamsAndGrads>(kParamsAndGrads);
-  if (!result.Has(kFusedGrads)) {
-    PADDLE_THROW(
-        "The alloc_continuous_space_for_grad_pass should be called before this "
-        "pass.");
-  }
-  auto &fused_grad = result.Get<FusedGrads>(kFusedGrads);
-  auto &fused_vars = result.Get<FusedVars>(kFusedVars);
-  auto iter = std::find(fused_vars.begin(), fused_vars.end(), fused_grad);
-  PADDLE_ENFORCE(iter != fused_vars.end(), "Not find the fused_grad.");
-  fused_vars_name.emplace("Grad", fused_grad);
-
-  // Step 4: Sort the parameters and auxiliary variables according
-  // to parameters' name to make variables' name correspond correctly.
-  PADDLE_ENFORCE(result.Has(kParamsAndGrads), "Does't find kParamsAndGrads.");
-  PADDLE_ENFORCE_EQ(params_grads.size(), aux_var_set.begin()->second.size(),
-                    "The size of params_grads and aux_var_set are not equal.");
-  SortParametersAndAuxVars(params_grads, &aux_var_set, &opt_ops);
-
-  // Step 5: Alloc continuous space for Parameters and AuxiliaryVar(e.g.
+  bool grad_fused = false;
+  if (result.Has(kParamsAndGrads)) {
+    auto &params_grads = result.Get<ParamsAndGrads>(kParamsAndGrads);
+    PADDLE_ENFORCE_EQ(
+        params_grads.size(), aux_var_set.at(kGrad).size(),
+        "The number of gradients and optimizer ops is not equal.");
+    std::unordered_set<std::string> opt_grad_set(aux_var_set.at(kGrad).begin(),
+                                                 aux_var_set.at(kGrad).end());
+    size_t same_grad_num = 0;
+    for (auto &p_g : params_grads) {
+      if (opt_grad_set.count(p_g.second)) {
+        ++same_grad_num;
+      }
+    }
+
+    // NOTE(zcd): the gradient of kParamsAndGrads may be different with the
+    // kGrad.
+    if (same_grad_num == aux_var_set.at(kGrad).size()) {
+      if (!result.Has(kFusedGrads)) {
+        PADDLE_THROW(
+            "The alloc_continuous_space_for_grad_pass should be called before "
+            "this pass.");
+      }
+      auto &fused_grad = result.Get<FusedGrads>(kFusedGrads);
+      auto &fused_vars = result.Get<FusedVars>(kFusedVars);
+      auto iter = std::find(fused_vars.begin(), fused_vars.end(), fused_grad);
+      PADDLE_ENFORCE(iter != fused_vars.end(), "Not find the fused_grad.");
+      fused_vars_name[kGrad] = fused_grad;
+
+      // Sort the parameters and auxiliary variables according
+      // to parameters' name to make variables' name correspond correctly.
+      SortParametersAndAuxVars(params_grads, &aux_var_set, &opt_ops);
+      grad_fused = true;
+    }
+  }
+
+  // Step 4: Alloc continuous space for Parameters and AuxiliaryVar(e.g.
   // Moment1, Moment2, Beta1Pow, Beta2Pow) of all the optimizer ops separately.
+  aux_var_names.pop_back();
+  if (!grad_fused) {
+    InitFusedGradsAndAllocSpaceForGrads(
+        places, local_scopes, aux_var_set.at(kParam), aux_var_set.at(kGrad),
+        fused_vars_name.at(kGrad), &result);
+  }
   InitFusedVarsAndAllocSpaceForVars(places, local_scopes, aux_var_names,
                                     aux_var_set, fused_vars_name);
 
-  // Step 6: Fuse optimizer Ops and Scale Ops
+  // Step 5: Fuse optimizer Ops and Scale Ops
   FuseOptimizerOps(aux_var_set, fused_vars_name, opt_ops, &result);
 
-  // Step 7: Remove optimizer Ops
+  // Step 6: Remove optimizer Ops
   for (auto &opt_op : opt_ops) {
     graph->RemoveNode(opt_op);
   }
 }
 
+void FuseOptimizerOpPass::InitFusedGradsAndAllocSpaceForGrads(
+    const std::vector<platform::Place> &places,
+    const std::vector<Scope *> &local_scopes,
+    const std::vector<std::string> &params,
+    const std::vector<std::string> &grads, const std::string &fused_grad_name,
+    ir::Graph *result) const {
+  // Get Var Nodes
+  std::unordered_map<std::string, ir::Node *> vars;
+  for (ir::Node *node : result->Nodes()) {
+    if (node->IsVar() && node->Var()) {
+      // Note: The graph may have the same name node. For example, parameter
+      // is the input of operator and it also is the output of optimizer;
+      vars.emplace(node->Var()->Name(), node);
+    }
+  }
+  // Init Grads
+  for (auto it = local_scopes.rbegin(); it != local_scopes.rend(); ++it) {
+    auto &scope = *it;
+    VLOG(10) << "Init " << fused_grad_name;
+    PADDLE_ENFORCE(scope->FindVar(fused_grad_name) == nullptr,
+                   "%s has existed in scope.", fused_grad_name);
+    scope->Var(fused_grad_name)->GetMutable<LoDTensor>();
+
+    for (auto &grad_var_name : grads) {
+      auto iter = vars.find(grad_var_name);
+      PADDLE_ENFORCE(iter != vars.end());
+      PADDLE_ENFORCE_NOT_NULL(iter->second->Var());
+      PADDLE_ENFORCE_EQ(iter->second->Var()->GetType(),
+                        proto::VarType::LOD_TENSOR);
+      scope->Var(grad_var_name)->GetMutable<LoDTensor>();
+    }
+  }
+  // Define Ops
+  ProgramDesc program_desc;
+  auto *global_block = program_desc.MutableBlock(0);
+  AppendAllocContinuousSpace(params, grads, fused_grad_name, global_block,
+                             false, false);
+  // Run Ops
+  RunInitOps(places, local_scopes, *global_block);
+}
+
 void FuseOptimizerOpPass::InitFusedVarsAndAllocSpaceForVars(
     const std::vector<platform::Place> &places,
     const std::vector<Scope *> &local_scopes,
@@ -115,37 +181,49 @@ void FuseOptimizerOpPass::InitFusedVarsAndAllocSpaceForVars(
     const std::unordered_map<std::string, std::vector<std::string>>
         &aux_var_set,
     const std::unordered_map<std::string, std::string> &fused_vars_name) const {
-  VLOG(10) << "Init FusedVars.";
-  // Alloc parameters and auxiliary vars in the respective scope.
-  size_t idx = local_scopes.size();
-  for (auto iter = local_scopes.rbegin(); iter != local_scopes.rend();
-       ++iter, --idx) {
-    auto &scope = *iter;
-    for (auto &var_name : aux_var_names) {
-      auto fused_var_name = fused_vars_name.at(var_name);
-      VLOG(10) << "Init " << fused_var_name;
-      PADDLE_ENFORCE(scope->FindVar(fused_var_name) == nullptr,
-                     "%s has exist in scope[%d]", fused_var_name, idx);
-      scope->Var(fused_var_name)->GetMutable<LoDTensor>();
-    }
+  // Init Vars
+  for (auto &var_name : aux_var_names) {
+    auto &fused_var_name = fused_vars_name.at(var_name);
+    InitVars(local_scopes, fused_var_name);
   }
-
+  // Define Ops
   ProgramDesc program_desc;
   auto *global_block = program_desc.MutableBlock(0);
   for (auto &var_name : aux_var_names) {
-    AppendAllocContinuousSpace(aux_var_set.at(var_name),
-                               fused_vars_name.at(var_name), true,
-                               global_block);
+    AppendAllocContinuousSpace(
+        aux_var_set.at(var_name), aux_var_set.at(var_name),
+        fused_vars_name.at(var_name), global_block, true);
   }
+  // Run Ops
+  RunInitOps(places, local_scopes, *global_block);
+}
 
+void FuseOptimizerOpPass::RunInitOps(const std::vector<platform::Place> &places,
+                                     const std::vector<Scope *> &local_scopes,
+                                     const BlockDesc &global_block) const {
   for (size_t i = 0; i < local_scopes.size(); ++i) {
-    for (auto &op_desc : global_block->AllOps()) {
+    for (auto &op_desc : global_block.AllOps()) {
       auto op = OpRegistry::CreateOp(*op_desc);
       op->Run(*local_scopes[i], places[i]);
     }
   }
 }
 
+void FuseOptimizerOpPass::InitVars(const std::vector<Scope *> &local_scopes,
+                                   const std::string &fused_var_name) const {
+  VLOG(10) << "Init FusedVars.";
+  // Alloc parameters and auxiliary vars in the respective scope.
+  size_t idx = local_scopes.size();
+  for (auto iter = local_scopes.rbegin(); iter != local_scopes.rend();
+       ++iter, --idx) {
+    auto &scope = *iter;
+    VLOG(10) << "Init " << fused_var_name;
+    PADDLE_ENFORCE(scope->FindVar(fused_var_name) == nullptr,
+                   "%s has exist in scope[%d]", fused_var_name, idx);
+    scope->Var(fused_var_name)->GetMutable<LoDTensor>();
+  }
+}
+
 void FuseOptimizerOpPass::SortParametersAndAuxVars(
     const std::vector<std::pair<std::string, std::string>> &params_grads,
     std::unordered_map<std::string, std::vector<std::string>> *aux_vars_set,
@@ -203,15 +281,16 @@ void FuseOptimizerOpPass::GetSpecifiedOpsAndVars(
 }
 
 void FuseOptimizerOpPass::AppendAllocContinuousSpace(
-    const std::vector<std::string> &args, const std::string &out_arg,
-    bool copy_data, BlockDesc *global_block) const {
+    const std::vector<std::string> &in_args,
+    const std::vector<std::string> &out_args, const std::string &fused_out_arg,
+    BlockDesc *global_block, bool copy_data, bool check_name) const {
   auto op_desc = global_block->AppendOp();
   op_desc->SetType("alloc_continuous_space");
-  op_desc->SetInput("Input", args);
-  op_desc->SetOutput("Output", args);
-  op_desc->SetOutput("FusedOutput", {out_arg});
+  op_desc->SetInput("Input", in_args);
+  op_desc->SetOutput("Output", out_args);
+  op_desc->SetOutput("FusedOutput", {fused_out_arg});
   op_desc->SetAttr("copy_data", copy_data);
-  op_desc->SetAttr("check_name", true);
+  op_desc->SetAttr("check_name", check_name);
 }
 
 void FuseOptimizerOpPass::InserInputAndOutputForOptOps(
diff --git a/paddle/fluid/framework/details/fuse_optimizer_op_pass.h b/paddle/fluid/framework/details/fuse_optimizer_op_pass.h
index 0240f1594d7ef9d855eb6e96e8e8a32ee1d957ba..47efc1693dd31ca88787da3a9d6d06aa7ef65786 100644
--- a/paddle/fluid/framework/details/fuse_optimizer_op_pass.h
+++ b/paddle/fluid/framework/details/fuse_optimizer_op_pass.h
@@ -27,6 +27,10 @@ namespace paddle {
 namespace framework {
 namespace details {
 
+constexpr char kGrad[] = "Grad";
+constexpr char kParam[] = "Param";
+constexpr char kLearningRate[] = "LearningRate";
+
 class FuseOptimizerOpPass : public ir::Pass {
  protected:
   void ApplyImpl(ir::Graph *graph) const override;
@@ -56,9 +60,18 @@ class FuseOptimizerOpPass : public ir::Pass {
       std::unordered_map<std::string, std::vector<std::string>> *aux_args_name)
       const;
 
-  void AppendAllocContinuousSpace(const std::vector<std::string> &args,
-                                  const std::string &out_arg, bool copy_data,
-                                  BlockDesc *global_block) const;
+  void AppendAllocContinuousSpace(const std::vector<std::string> &in_args,
+                                  const std::vector<std::string> &out_args,
+                                  const std::string &fused_out_arg,
+                                  BlockDesc *global_block, bool copy_data,
+                                  bool check_name = true) const;
+
+  void InitFusedGradsAndAllocSpaceForGrads(
+      const std::vector<platform::Place> &places,
+      const std::vector<Scope *> &local_scopes,
+      const std::vector<std::string> &params,
+      const std::vector<std::string> &grads, const std::string &fused_grad_name,
+      ir::Graph *result) const;
 
   void InitFusedVarsAndAllocSpaceForVars(
       const std::vector<platform::Place> &places,
@@ -68,6 +81,13 @@ class FuseOptimizerOpPass : public ir::Pass {
           &aux_var_set,
       const std::unordered_map<std::string, std::string> &fused_vars_name)
       const;
+
+  void RunInitOps(const std::vector<platform::Place> &places,
+                  const std::vector<Scope *> &local_scopes,
+                  const BlockDesc &global_block) const;
+
+  void InitVars(const std::vector<Scope *> &local_scopes,
+                const std::string &fused_var_name) const;
 };
 
 }  // namespace details
diff --git a/paddle/fluid/framework/details/fuse_sgd_op_pass.cc b/paddle/fluid/framework/details/fuse_sgd_op_pass.cc
index f91c21e3cc869de1a6d67146eb99f27a2ca5497c..2219f3209f77de5cb34abfb9edb8bdea6a8eebb0 100644
--- a/paddle/fluid/framework/details/fuse_sgd_op_pass.cc
+++ b/paddle/fluid/framework/details/fuse_sgd_op_pass.cc
@@ -24,7 +24,7 @@ namespace details {
 const std::string FuseSgdOpPass::GetOpType() const { return "sgd"; }
 
 const std::vector<std::string> FuseSgdOpPass::GetAuxiliaryVarNames() const {
-  return {"Param"};
+  return {};
 }
 
 void FuseSgdOpPass::FuseOptimizerOps(
@@ -50,12 +50,12 @@ void FuseSgdOpPass::FuseSgdOps(
   // Add fused scale
   OpDesc Sgd_desc(sgd_ops[0]->Op()->Block());
   Sgd_desc.SetType("sgd");
-  Sgd_desc.SetInput("Param", {fused_vars_name.at("Param")});
-  Sgd_desc.SetInput("Grad", {fused_vars_name.at("Grad")});
-  Sgd_desc.SetOutput("ParamOut", {fused_vars_name.at("Param")});
+  Sgd_desc.SetInput(kParam, {fused_vars_name.at(kParam)});
+  Sgd_desc.SetInput(kGrad, {fused_vars_name.at(kGrad)});
+  Sgd_desc.SetOutput("ParamOut", {fused_vars_name.at(kParam)});
 
   // TODO(zcd): The LearningRate, Beta1Pow, Beta2Pow should be equal.
-  Sgd_desc.SetInput("LearningRate", sgd_ops[0]->Op()->Input("LearningRate"));
+  Sgd_desc.SetInput(kLearningRate, sgd_ops[0]->Op()->Input(kLearningRate));
 
   // NOTE: multi_devices_pass requires that every op should have a role.
   Sgd_desc.SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(), op_role);
diff --git a/paddle/fluid/framework/details/parallel_ssa_graph_executor.cc b/paddle/fluid/framework/details/parallel_ssa_graph_executor.cc
index 137e0dd7708dcc77c3a927979cfb357249f1fdc9..1bd27263f7dad5f733c553c202444ba7cacd2510 100644
--- a/paddle/fluid/framework/details/parallel_ssa_graph_executor.cc
+++ b/paddle/fluid/framework/details/parallel_ssa_graph_executor.cc
@@ -106,7 +106,7 @@ ParallelSSAGraphExecutor::ParallelSSAGraphExecutor(
   VLOG(1) << "set num_threads: " << strategy_.num_threads_
           << " to run the operators of the graph on each device.";
   for (size_t i = 0; i < places.size(); ++i) {
-    executors_.emplace_back(new details::ThreadedSSAGraphExecutor(
+    executors_.emplace_back(new details::FastThreadedSSAGraphExecutor(
         strategy_, local_scopes_, {places_[i]}, graphs_.at(i).get()));
   }
 }
diff --git a/paddle/fluid/framework/details/parallel_ssa_graph_executor.h b/paddle/fluid/framework/details/parallel_ssa_graph_executor.h
index 1e421f2a3a51363fe368859f7a34593c8c894077..faf071b05306a49c0049421bc72e4981c0bfc84c 100644
--- a/paddle/fluid/framework/details/parallel_ssa_graph_executor.h
+++ b/paddle/fluid/framework/details/parallel_ssa_graph_executor.h
@@ -14,12 +14,12 @@
 
 #pragma once
 
+#include <memory>
 #include <string>
 #include <vector>
-
 #include "ThreadPool.h"
+#include "paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h"
 #include "paddle/fluid/framework/details/multi_devices_helper.h"
-#include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
 #include "paddle/fluid/framework/ir/graph.h"
 
 namespace paddle {
@@ -48,7 +48,8 @@ class ParallelSSAGraphExecutor : public SSAGraphExecutor {
   std::vector<platform::Place> places_;
   std::vector<std::unique_ptr<ir::Graph>> graphs_;
 
-  std::vector<std::unique_ptr<details::ThreadedSSAGraphExecutor>> executors_;
+  std::vector<std::unique_ptr<details::FastThreadedSSAGraphExecutor>>
+      executors_;
   ExceptionHolder exception_holder_;
 };
 
diff --git a/paddle/fluid/framework/downpour_worker.cc b/paddle/fluid/framework/downpour_worker.cc
index 4ca7842fa261a1b8178438d35ca5d626146663d4..386ffd84c57063e950cd8b0d57304c66190be4c4 100644
--- a/paddle/fluid/framework/downpour_worker.cc
+++ b/paddle/fluid/framework/downpour_worker.cc
@@ -21,40 +21,40 @@ namespace framework {
 
 void DownpourWorker::Initialize(const TrainerDesc& desc) {
   param_ = desc.downpour_param();
-  for (size_t i = 0; i < param_.sparse_table_size(); ++i) {
+  for (int i = 0; i < param_.sparse_table_size(); ++i) {
     uint64_t table_id =
         static_cast<uint64_t>(param_.sparse_table(i).table_id());
     TableParameter table = param_.sparse_table(i);
     sparse_key_names_[table_id].resize(table.sparse_key_name_size());
-    for (size_t j = 0; j < table.sparse_key_name_size(); ++j) {
+    for (int j = 0; j < table.sparse_key_name_size(); ++j) {
       sparse_key_names_[table_id][j] = table.sparse_key_name(j);
     }
     sparse_value_names_[table_id].resize(table.sparse_value_name_size());
-    for (size_t j = 0; j < table.sparse_value_name_size(); ++j) {
+    for (int j = 0; j < table.sparse_value_name_size(); ++j) {
       sparse_value_names_[table_id][j] = table.sparse_value_name(j);
     }
     sparse_grad_names_[table_id].resize(table.sparse_grad_name_size());
-    for (size_t j = 0; j < table.sparse_grad_name_size(); ++j) {
+    for (int j = 0; j < table.sparse_grad_name_size(); ++j) {
       sparse_grad_names_[table_id][j] = table.sparse_grad_name(j);
     }
     label_var_name_[table_id] = table.label_var_name();
   }
 
-  for (size_t i = 0; i < param_.dense_table_size(); ++i) {
+  for (int i = 0; i < param_.dense_table_size(); ++i) {
     uint64_t table_id = static_cast<uint64_t>(param_.dense_table(i).table_id());
     auto table = param_.dense_table(i);
     dense_value_names_[table_id].resize(table.dense_value_name_size());
-    for (size_t j = 0; j < table.dense_value_name_size(); ++j) {
+    for (int j = 0; j < table.dense_value_name_size(); ++j) {
       dense_value_names_[table_id][j] = table.dense_value_name(j);
     }
     dense_grad_names_[table_id].resize(table.dense_grad_name_size());
-    for (size_t j = 0; j < table.dense_grad_name_size(); ++j) {
+    for (int j = 0; j < table.dense_grad_name_size(); ++j) {
       dense_grad_names_[table_id][j] = table.dense_grad_name(j);
     }
   }
 
   skip_ops_.resize(param_.skip_ops_size());
-  for (size_t i = 0; i < param_.skip_ops_size(); ++i) {
+  for (int i = 0; i < param_.skip_ops_size(); ++i) {
     skip_ops_[i] = param_.skip_ops(i);
   }
 
@@ -83,14 +83,14 @@ void DownpourWorker::CollectLabelInfo(size_t table_idx) {
   LoDTensor* tensor = var->GetMutable<LoDTensor>();
   int64_t* label_ptr = tensor->data<int64_t>();
 
-  int global_index = 0;
+  size_t global_index = 0;
   for (size_t i = 0; i < sparse_key_names_[table_id].size(); ++i) {
     VLOG(3) << "sparse_key_names_[" << i
             << "]: " << sparse_key_names_[table_id][i];
     Variable* fea_var = thread_scope_->FindVar(sparse_key_names_[table_id][i]);
     LoDTensor* tensor = fea_var->GetMutable<LoDTensor>();
     int64_t* ids = tensor->data<int64_t>();
-    int fea_idx = 0;
+    size_t fea_idx = 0;
     // tensor->lod()[0].size() == batch_size + 1
     for (auto lod_idx = 1u; lod_idx < tensor->lod()[0].size(); ++lod_idx) {
       for (; fea_idx < tensor->lod()[0][lod_idx]; ++fea_idx) {
@@ -138,7 +138,7 @@ void DownpourWorker::FillSparseValue(size_t table_idx) {
     auto& tensor_lod = tensor->lod()[0];
     LoD data_lod{tensor_lod};
     tensor_emb->set_lod(data_lod);
-    for (auto index = 0u; index < len; ++index) {
+    for (int index = 0; index < len; ++index) {
       if (ids[index] == 0u) {
         memcpy(ptr + table.emb_dim() * index, init_value.data() + 2,
                sizeof(float) * table.emb_dim());
@@ -192,7 +192,7 @@ void DownpourWorker::TrainFilesWithProfiler() {
     read_time += timeline.ElapsedSec();
     total_time += timeline.ElapsedSec();
     VLOG(3) << "program config size: " << param_.program_config_size();
-    for (size_t i = 0; i < param_.program_config(0).pull_sparse_table_id_size();
+    for (int i = 0; i < param_.program_config(0).pull_sparse_table_id_size();
          ++i) {
       uint64_t tid = static_cast<uint64_t>(
           param_.program_config(0).pull_sparse_table_id(i));
@@ -244,8 +244,8 @@ void DownpourWorker::TrainFilesWithProfiler() {
     }
 
     if (need_to_push_sparse_) {
-      for (size_t i = 0;
-           i < param_.program_config(0).push_sparse_table_id_size(); ++i) {
+      for (int i = 0; i < param_.program_config(0).push_sparse_table_id_size();
+           ++i) {
         uint64_t tid = static_cast<uint64_t>(
             param_.program_config(0).push_sparse_table_id(i));
         TableParameter table;
@@ -268,8 +268,8 @@ void DownpourWorker::TrainFilesWithProfiler() {
 
     if (need_to_push_dense_) {
       timeline.Start();
-      for (size_t i = 0;
-           i < param_.program_config(0).push_dense_table_id_size(); ++i) {
+      for (int i = 0; i < param_.program_config(0).push_dense_table_id_size();
+           ++i) {
         uint64_t tid = static_cast<uint64_t>(
             param_.program_config(0).push_dense_table_id(i));
         fleet_ptr_->PushDenseVarsAsync(
@@ -315,8 +315,8 @@ void DownpourWorker::TrainFilesWithProfiler() {
     }
 
     if (need_to_push_dense_) {
-      for (size_t i = 0;
-           i < param_.program_config(0).push_dense_table_id_size(); ++i) {
+      for (int i = 0; i < param_.program_config(0).push_dense_table_id_size();
+           ++i) {
         uint64_t tid = static_cast<uint64_t>(
             param_.program_config(0).push_dense_table_id(i));
         pull_dense_worker_->IncreaseThreadVersion(thread_id_, tid);
@@ -362,7 +362,7 @@ void DownpourWorker::TrainFiles() {
   int cur_batch;
   while ((cur_batch = device_reader_->Next()) > 0) {
     // pull sparse here
-    for (size_t i = 0; i < param_.program_config(0).pull_sparse_table_id_size();
+    for (int i = 0; i < param_.program_config(0).pull_sparse_table_id_size();
          ++i) {
       uint64_t tid = static_cast<uint64_t>(
           param_.program_config(0).pull_sparse_table_id(i));
@@ -397,8 +397,8 @@ void DownpourWorker::TrainFiles() {
 
     if (need_to_push_sparse_) {
       // push gradients here
-      for (size_t i = 0;
-           i < param_.program_config(0).push_sparse_table_id_size(); ++i) {
+      for (int i = 0; i < param_.program_config(0).push_sparse_table_id_size();
+           ++i) {
         uint64_t tid = static_cast<uint64_t>(
             param_.program_config(0).push_sparse_table_id(i));
         TableParameter table;
@@ -416,8 +416,8 @@ void DownpourWorker::TrainFiles() {
     }
 
     if (need_to_push_dense_) {
-      for (size_t i = 0;
-           i < param_.program_config(0).push_dense_table_id_size(); ++i) {
+      for (int i = 0; i < param_.program_config(0).push_dense_table_id_size();
+           ++i) {
         uint64_t tid = static_cast<uint64_t>(
             param_.program_config(0).push_dense_table_id(i));
         fleet_ptr_->PushDenseVarsAsync(
@@ -461,8 +461,8 @@ void DownpourWorker::TrainFiles() {
     }
 
     if (need_to_push_dense_) {
-      for (size_t i = 0;
-           i < param_.program_config(0).push_dense_table_id_size(); ++i) {
+      for (int i = 0; i < param_.program_config(0).push_dense_table_id_size();
+           ++i) {
         uint64_t tid = static_cast<uint64_t>(
             param_.program_config(0).push_dense_table_id(i));
         pull_dense_worker_->IncreaseThreadVersion(thread_id_, tid);
diff --git a/paddle/fluid/framework/ir/expected_kernel_cache_pass.cc b/paddle/fluid/framework/ir/expected_kernel_cache_pass.cc
index ee67af0aff5c90a9da0ece8f197d9a0c0a8e5b9c..4a99d4c1a9c0f0bd973097d281e380341fe88515 100644
--- a/paddle/fluid/framework/ir/expected_kernel_cache_pass.cc
+++ b/paddle/fluid/framework/ir/expected_kernel_cache_pass.cc
@@ -23,7 +23,7 @@ namespace ir {
 void ExpectedKernelCachePass::ApplyImpl(ir::Graph* graph) const {
   VLOG(3) << "Applies Expected Kernel Cache strategy.";
   for (const Node* n : graph->Nodes()) {
-    if (n->IsOp()) {
+    if (n->IsOp() && n->Op()) {
       n->Op()->SetAttr(kEnableCacheExpectedKernel, true);
     }
   }
diff --git a/paddle/fluid/framework/ir/graph_helper.cc b/paddle/fluid/framework/ir/graph_helper.cc
index 28a37f331c100695f0ffec7288db84f4493d68a0..12ce99c8788625e2aae6e07abdea565bb2c2ebb9 100644
--- a/paddle/fluid/framework/ir/graph_helper.cc
+++ b/paddle/fluid/framework/ir/graph_helper.cc
@@ -31,10 +31,10 @@ namespace paddle {
 namespace framework {
 namespace ir {
 namespace {
-void SortHelper(
-    const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list,
-    ir::Node *node, std::unordered_set<ir::Node *> *visited,
-    std::vector<ir::Node *> *ret) {
+void SortHelper(const std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>,
+                               ir::NodeComp> &adj_list,
+                ir::Node *node, std::unordered_set<ir::Node *> *visited,
+                std::vector<ir::Node *> *ret) {
   visited->insert(node);
 
   for (auto adj : adj_list.at(node)) {
@@ -50,7 +50,8 @@ void SortHelper(
 
 bool HasCircleHelper(
     ir::Node *node,
-    const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list,
+    const std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
+        &adj_list,
     std::unordered_set<ir::Node *> *visited,
     std::unordered_set<ir::Node *> *in_trace,
     std::vector<std::vector<ir::Node *>> *circles) {
@@ -84,7 +85,8 @@ bool HasCircleHelper(
 }
 
 bool HasCircleInternal(
-    const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list,
+    const std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
+        &adj_list,
     std::vector<std::vector<ir::Node *>> *circles) {
   std::unordered_set<ir::Node *> visited;
   std::unordered_set<ir::Node *> in_trace;
@@ -107,8 +109,8 @@ bool FindCircleSubGraph(const Graph &graph,
 }
 
 std::vector<ir::Node *> TopologySortOperations(const Graph &graph) {
-  std::map<ir::Node *, std::unordered_set<ir::Node *>> adj_list =
-      BuildOperationAdjList(graph);
+  std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
+      adj_list = BuildOperationAdjList(graph);
   PADDLE_ENFORCE(!HasCircleInternal(adj_list, nullptr));
   std::unordered_set<ir::Node *> visited;
   std::vector<ir::Node *> ret;
@@ -117,34 +119,30 @@ std::vector<ir::Node *> TopologySortOperations(const Graph &graph) {
       SortHelper(adj_list, adj.first, &visited, &ret);
     }
   }
+
   return ret;
 }
 
 // Build operator inlink edge table.
-std::map<ir::Node *, std::unordered_set<ir::Node *>> BuildOperationAdjList(
-    const Graph &graph) {
-  std::map<ir::Node *, std::unordered_set<ir::Node *>> adj_list;
+std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
+BuildOperationAdjList(const Graph &graph) {
+  std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
+      adj_list;
 
   for (auto &n : graph.Nodes()) {
     if (!n->IsOp()) continue;
     if (adj_list.find(n) == adj_list.end()) {
-      adj_list[n] = std::unordered_set<ir::Node *>();
+      adj_list[n] = std::set<ir::Node *, ir::NodeComp>();
     }
-    std::vector<ir::Node *> nodes;
     for (auto &var : n->inputs) {
       for (auto &adj_n : var->inputs) {
         PADDLE_ENFORCE(adj_n->NodeType() == ir::Node::Type::kOperation);
         VLOG(4) << "adj " << adj_n->Name() << reinterpret_cast<void *>(adj_n)
                 << " -> " << n->Name() << reinterpret_cast<void *>(n)
                 << "  via " << var->Name() << reinterpret_cast<void *>(var);
-        nodes.push_back(adj_n);
+        adj_list[n].insert(adj_n);
       }
     }
-    std::sort(nodes.begin(), nodes.end(), [](ir::Node *node1, ir::Node *node2) {
-      return node1->id() > node2->id();
-    });
-    adj_list[n].insert(std::make_move_iterator(nodes.begin()),
-                       std::make_move_iterator(nodes.end()));
   }
   return adj_list;
 }
diff --git a/paddle/fluid/framework/ir/graph_helper.h b/paddle/fluid/framework/ir/graph_helper.h
index 214de9ec7d85aee6021b18866295777e317aa79d..849a9c3be6904f3f9c3669d8fc9d750154863031 100644
--- a/paddle/fluid/framework/ir/graph_helper.h
+++ b/paddle/fluid/framework/ir/graph_helper.h
@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include <map>
 #include <memory>
+#include <set>
 #include <vector>
 
 #include "paddle/fluid/framework/ir/graph.h"
@@ -25,6 +26,13 @@ namespace paddle {
 namespace framework {
 namespace ir {
 
+// Compare nodes via node id.
+struct NodeComp {
+  bool operator()(ir::Node *const &node1, ir::Node *const &node2) const {
+    return node1->id() < node2->id();
+  }
+};
+
 // Test if the graph contains circle.
 bool HasCircle(const Graph &graph);
 
@@ -57,8 +65,8 @@ std::vector<Node *> TopologyVarientSort(const Graph &graph, SortKind sort_kind);
 void CleanIndividualNodes(Graph *graph);
 
 // Build an adjacency list of operations for the `graph`.
-std::map<ir::Node *, std::unordered_set<ir::Node *>> BuildOperationAdjList(
-    const Graph &graph);
+std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
+BuildOperationAdjList(const Graph &graph);
 
 template <typename T>
 std::vector<T *> FilterByNodeWrapper(const Graph &graph) {
diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc
index e6f5b15af8cd440a9304235acfe62787c5f1b134..1ea93b7638a85e67bcc85a0c0e130d636938d6c5 100644
--- a/paddle/fluid/framework/op_desc.cc
+++ b/paddle/fluid/framework/op_desc.cc
@@ -241,6 +241,7 @@ OpDesc::OpDesc(const std::string &type, const VariableNameMap &inputs,
   outputs_ = outputs;
   attrs_ = attrs;
   need_update_ = true;
+  block_ = nullptr;
 }
 
 OpDesc::OpDesc(const OpDesc &other, BlockDesc *block) {
diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc
index 4245caf1689c76d72b410c742488c55562c8b998..c4bf2b7e8c017b22f917c9f9bd40e75b8cde08b2 100644
--- a/paddle/fluid/framework/parallel_executor.cc
+++ b/paddle/fluid/framework/parallel_executor.cc
@@ -221,7 +221,7 @@ ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
     PADDLE_ENFORCE(!member_->use_cuda_,
                    "gpu mode does not support async_mode_ now!");
     graphs.push_back(graph);
-    for (int i = 1; i < places.size(); ++i) {
+    for (size_t i = 1; i < places.size(); ++i) {
       auto *tmp_graph = new ir::Graph(graph->OriginProgram());
       async_graphs_.emplace_back(tmp_graph);
       graphs.push_back(tmp_graph);
@@ -315,7 +315,7 @@ ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
     graph = build_strategy.Apply(graph, {member_->places_[0]}, loss_var_name,
                                  {member_->local_scopes_[0]}, 1,
                                  member_->use_cuda_, member_->nccl_ctxs_.get());
-    for (int i = 1; i < member_->places_.size(); ++i) {
+    for (size_t i = 1; i < member_->places_.size(); ++i) {
       graphs[i] =
           build_strategy.Apply(graphs[i], {member_->places_[i]}, loss_var_name,
                                {member_->local_scopes_[i]}, 1,
diff --git a/paddle/fluid/framework/trainer_desc.proto b/paddle/fluid/framework/trainer_desc.proto
index 389c1a870fb54ad28806ad49632323b1c93676f4..4fc05ccf5c9be37e80b4ae7263166ad76eb6d6a7 100644
--- a/paddle/fluid/framework/trainer_desc.proto
+++ b/paddle/fluid/framework/trainer_desc.proto
@@ -76,7 +76,7 @@ message PullDenseWorkerParameter {
 
 message TableParameter {
   // dense table only
-  optional int64 table_id = 1;
+  optional uint64 table_id = 1;
   repeated string dense_value_name = 2;
   repeated string dense_grad_name = 3;
   repeated int32 push_dense_wait_times = 5;
diff --git a/paddle/fluid/framework/var_type_inference.h b/paddle/fluid/framework/var_type_inference.h
index 2e9c64d3e6854bf70c0aee06128b9f1b7c8c7439..66e6ac81623a1cd1c79981c1e4a97d974e9c2426 100644
--- a/paddle/fluid/framework/var_type_inference.h
+++ b/paddle/fluid/framework/var_type_inference.h
@@ -45,12 +45,16 @@ class InferVarTypeContext {
 
   virtual bool HasInput(const std::string& name) const {
     PADDLE_ENFORCE_NOT_NULL(op_);
-    return op_->Inputs().count(name) > 0;
+    auto& inputs = op_->Inputs();
+    auto input = inputs.find(name);
+    return input != inputs.end() && !input->second.empty();
   }
 
   virtual bool HasOutput(const std::string& name) const {
     PADDLE_ENFORCE_NOT_NULL(op_);
-    return op_->Outputs().count(name) > 0;
+    auto& outputs = op_->Outputs();
+    auto output = outputs.find(name);
+    return output != outputs.end() && !output->second.empty();
   }
 
   virtual const std::vector<std::string>& Input(const std::string& name) const {
diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc
index 6942604b0723f8665f0e8b058d48a5356a1a01f4..0155609a029664da2c3d4c90a152ec556927c32d 100644
--- a/paddle/fluid/inference/api/analysis_predictor.cc
+++ b/paddle/fluid/inference/api/analysis_predictor.cc
@@ -259,6 +259,9 @@ bool AnalysisPredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
       return false;
     }
 
+    PADDLE_ENFORCE_NOT_NULL(input_ptr);
+    PADDLE_ENFORCE_NOT_NULL(inputs[i].data.data());
+
     if (platform::is_cpu_place(place_)) {
       // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
       std::memcpy(static_cast<void *>(input_ptr), inputs[i].data.data(),
diff --git a/paddle/fluid/inference/api/api.cc b/paddle/fluid/inference/api/api.cc
index 7d57b6ec74468dbdb0519f85140629a0ac01c18d..fc2d7b48c2a1f89232dcb96d1899667230e2ddda 100644
--- a/paddle/fluid/inference/api/api.cc
+++ b/paddle/fluid/inference/api/api.cc
@@ -54,6 +54,7 @@ PaddleBuf &PaddleBuf::operator=(const PaddleBuf &other) {
     memory_owned_ = other.memory_owned_;
   } else {
     Resize(other.length());
+    PADDLE_ENFORCE(!(other.length() > 0 && other.data() == nullptr));
     memcpy(data_, other.data(), other.length());
     length_ = other.length();
     memory_owned_ = true;
diff --git a/paddle/fluid/inference/api/api_impl.cc b/paddle/fluid/inference/api/api_impl.cc
index 54f40563c3662af24e794422be4d3262d86c76a7..56996c5cff88f5b4a9094291a09996f8b8d70a23 100644
--- a/paddle/fluid/inference/api/api_impl.cc
+++ b/paddle/fluid/inference/api/api_impl.cc
@@ -169,6 +169,7 @@ std::unique_ptr<PaddlePredictor> NativePaddlePredictor::Clone() {
   std::unique_ptr<PaddlePredictor> cls(new NativePaddlePredictor(config_));
   // Hot fix the bug that result diff in multi-thread.
   // TODO(Superjomn) re-implement a real clone here.
+  PADDLE_ENFORCE_NOT_NULL(dynamic_cast<NativePaddlePredictor *>(cls.get()));
   if (!dynamic_cast<NativePaddlePredictor *>(cls.get())->Init(nullptr)) {
     LOG(ERROR) << "fail to call Init";
     return nullptr;
@@ -210,6 +211,8 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
       return false;
     }
 
+    PADDLE_ENFORCE_NOT_NULL(input_ptr);
+    PADDLE_ENFORCE_NOT_NULL(inputs[i].data.data());
     if (platform::is_cpu_place(place_)) {
       // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
       std::memcpy(static_cast<void *>(input_ptr), inputs[i].data.data(),
@@ -316,6 +319,8 @@ std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
   }
 
   std::unique_ptr<PaddlePredictor> predictor(new NativePaddlePredictor(config));
+  PADDLE_ENFORCE_NOT_NULL(
+      dynamic_cast<NativePaddlePredictor *>(predictor.get()));
   if (!dynamic_cast<NativePaddlePredictor *>(predictor.get())->Init(nullptr)) {
     return nullptr;
   }
diff --git a/paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc b/paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc
index 9f23b9f037bcaeb758312d011067ae29c82e73cd..5ee848c3cfa2117b2adeab5e563c5d07ce1d76ca 100644
--- a/paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc
+++ b/paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc
@@ -47,6 +47,7 @@ struct DataRecord {
       num_lines++;
       std::vector<std::string> data;
       split(line, '\t', &data);
+      PADDLE_ENFORCE(data.size() >= 4);
       // load title1 data
       std::vector<int64_t> title1_data;
       split_to_int64(data[0], ' ', &title1_data);
diff --git a/paddle/fluid/inference/tests/api/analyzer_transformer_tester.cc b/paddle/fluid/inference/tests/api/analyzer_transformer_tester.cc
index bd4f1b61973fb0de06dcc288e329c94756d5ed47..a23297f29cf65d891f530850ffd184aa58e10886 100644
--- a/paddle/fluid/inference/tests/api/analyzer_transformer_tester.cc
+++ b/paddle/fluid/inference/tests/api/analyzer_transformer_tester.cc
@@ -214,28 +214,23 @@ TEST(Analyzer_Transformer, fuse_statis) {
 }
 
 // Compare result of NativeConfig and AnalysisConfig
-// void compare(bool use_mkldnn = false) {
-//   AnalysisConfig cfg;
-//   SetConfig(&cfg);
-//   if (use_mkldnn) {
-//     cfg.EnableMKLDNN();
-//   }
-//
-//   std::vector<std::vector<PaddleTensor>> input_slots_all;
-//   SetInput(&input_slots_all);
-//   CompareNativeAndAnalysis(
-//       reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
-//       input_slots_all);
-// }
-
-// TODO(yihuaxu):
-//    Disable compare and compare_mkldnn temporary, see
-//    https://github.com/paddlePaddle/Paddle/issues/16316 for details.
-// TEST(Analyzer_Transformer, compare) { compare(); }
-// #ifdef PADDLE_WITH_MKLDNN
-// TEST(Analyzer_Transformer, compare_mkldnn) { compare(true /* use_mkldnn */);
-// }
-// #endif
+void compare(bool use_mkldnn = false) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+  if (use_mkldnn) {
+    cfg.EnableMKLDNN();
+  }
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  SetInput(&input_slots_all);
+  CompareNativeAndAnalysis(
+      reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
+}
+
+TEST(Analyzer_Transformer, compare) { compare(); }
+#ifdef PADDLE_WITH_MKLDNN
+TEST(Analyzer_Transformer, compare_mkldnn) { compare(true /* use_mkldnn */); }
+#endif
 
 }  // namespace inference
 }  // namespace paddle
diff --git a/paddle/fluid/op_use_default_grad_op_maker.spec b/paddle/fluid/op_use_default_grad_op_maker.spec
index 21a25ce7d5e2bad172cf50cee6138ef4b44b07c1..63eaa676a43fc784dce2437ca15bc85e2295dbb7 100644
--- a/paddle/fluid/op_use_default_grad_op_maker.spec
+++ b/paddle/fluid/op_use_default_grad_op_maker.spec
@@ -29,6 +29,8 @@ pool3d
 prelu
 quantize
 rank_loss
+reduce_all
+reduce_any
 reduce_max
 reduce_mean
 reduce_min
diff --git a/paddle/fluid/operators/detection/gpc.cc b/paddle/fluid/operators/detection/gpc.cc
index 7c0823c0487d39eece5be08322e7d182b931ba3c..f46aaf7d0a7b2d48f18ba6cccb555bbb691ad353 100644
--- a/paddle/fluid/operators/detection/gpc.cc
+++ b/paddle/fluid/operators/detection/gpc.cc
@@ -24,6 +24,7 @@
  **/
 
 #include "paddle/fluid/operators/detection/gpc.h"
+#include "paddle/fluid/platform/enforce.h"
 
 namespace gpc {
 
@@ -689,6 +690,7 @@ static bbox *create_contour_bboxes(gpc_polygon *p) {
 
   gpc_malloc<bbox>(box, p->num_contours * sizeof(bbox),
                    const_cast<char *>("Bounding box creation"));
+  PADDLE_ENFORCE_NOT_NULL(box);
 
   /* Construct contour bounding boxes */
   for (c = 0; c < p->num_contours; c++) {
@@ -852,6 +854,7 @@ void gpc_add_contour(gpc_polygon *p, gpc_vertex_list *new_contour, int hole) {
   /* Create an extended hole array */
   gpc_malloc<int>(extended_hole, (p->num_contours + 1) * sizeof(int),
                   const_cast<char *>("contour hole addition"));
+  PADDLE_ENFORCE_NOT_NULL(extended_hole);
 
   /* Create an extended contour array */
   gpc_malloc<gpc_vertex_list>(extended_contour,
@@ -969,6 +972,7 @@ void gpc_polygon_clip(gpc_op op, gpc_polygon *subj, gpc_polygon *clip,
   /* Build scanbeam table from scanbeam tree */
   gpc_malloc<double>(sbt, sbt_entries * sizeof(double),
                      const_cast<char *>("sbt creation"));
+  PADDLE_ENFORCE_NOT_NULL(sbt);
   build_sbt(&scanbeam, sbt, sbtree);
   scanbeam = 0;
   free_sbtree(&sbtree);
@@ -1604,6 +1608,7 @@ void gpc_tristrip_clip(gpc_op op, gpc_polygon *subj, gpc_polygon *clip,
   /* Build scanbeam table from scanbeam tree */
   gpc_malloc<double>(sbt, sbt_entries * sizeof(double),
                      const_cast<char *>("sbt creation"));
+  PADDLE_ENFORCE_NOT_NULL(sbt);
   build_sbt(&scanbeam, sbt, sbtree);
   scanbeam = 0;
   free_sbtree(&sbtree);
diff --git a/paddle/fluid/operators/distributed/CMakeLists.txt b/paddle/fluid/operators/distributed/CMakeLists.txt
index 972b4f67a8388ce68952fa90aaa224cd45c6d226..f6531ec9edca7b425d28853f542d5e46783ba699 100644
--- a/paddle/fluid/operators/distributed/CMakeLists.txt
+++ b/paddle/fluid/operators/distributed/CMakeLists.txt
@@ -9,6 +9,9 @@ else()
 endif()
 configure_file(send_recv.proto.in ${CMAKE_CURRENT_SOURCE_DIR}/send_recv.proto @ONLY)
 
+cc_library(async_sparse_param_update_recorder SRCS async_sparse_param_update_recorder.cc DEPS enforce simple_threadpool)
+cc_test(async_sparse_param_update_recorder_test SRCS async_sparse_param_update_recorder_test.cc DEPS async_sparse_param_update_recorder)
+
 # FIXME(typhoonzero): use add_subdirectory once we clean the dependency of these files
 set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
 if(WITH_GRPC)
@@ -20,7 +23,7 @@ if(WITH_GRPC)
         collective_client.cc collective_server.cc
         ${GRPC_SRCS}
       PROTO send_recv.proto 
-      DEPS lod_tensor selected_rows_functor memory scope ${GRPC_DEPS})
+      DEPS lod_tensor selected_rows_functor memory scope ${GRPC_DEPS} async_sparse_param_update_recorder)
 
   set_source_files_properties(grpc_serde_test.cc rpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
   set(RPC_DEPS sendrecvop_rpc ${GRPC_DEPS})
diff --git a/paddle/fluid/operators/distributed/async_sparse_param_update_recorder.cc b/paddle/fluid/operators/distributed/async_sparse_param_update_recorder.cc
new file mode 100644
index 0000000000000000000000000000000000000000..3f3b6b959e30194c10b1a58d6fc3e7a61ad01313
--- /dev/null
+++ b/paddle/fluid/operators/distributed/async_sparse_param_update_recorder.cc
@@ -0,0 +1,27 @@
+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/operators/distributed/async_sparse_param_update_recorder.h"
+
+namespace paddle {
+namespace operators {
+namespace distributed {
+
+std::once_flag AsyncSparseParamUpdateRecorder::init_flag_;
+std::unique_ptr<AsyncSparseParamUpdateRecorder>
+    AsyncSparseParamUpdateRecorder::recorder_(nullptr);
+
+}  // namespace distributed
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/fluid/operators/distributed/async_sparse_param_update_recorder.h b/paddle/fluid/operators/distributed/async_sparse_param_update_recorder.h
new file mode 100644
index 0000000000000000000000000000000000000000..eadd842c7f6ead56006fd0c34814b1b7bd9b62f4
--- /dev/null
+++ b/paddle/fluid/operators/distributed/async_sparse_param_update_recorder.h
@@ -0,0 +1,183 @@
+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <functional>
+#include <future>  // NOLINT
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+
+#include <ThreadPool.h>
+
+#include "paddle/fluid/platform/enforce.h"
+
+namespace paddle {
+namespace operators {
+namespace distributed {
+
+class ConcurrentSet {
+ public:
+  ConcurrentSet() : pool_(new ::ThreadPool(1)) {}
+  ~ConcurrentSet() {}
+
+  std::future<void> Update(const std::vector<int64_t>& rows) {
+    auto task = [this, rows] {
+      if (VLOG_IS_ON(3)) {
+        std::ostringstream sstream;
+        sstream << "[";
+        for (auto& id : rows) {
+          sstream << id << ", ";
+        }
+        sstream << "]";
+        VLOG(3) << "update ids -> " << sstream.str();
+      }
+      for (auto row : rows) {
+        set_.insert(row);
+      }
+    };
+    return pool_->enqueue(std::move(task));
+  }
+
+  std::future<void> GetAndClear(std::vector<int64_t>* result) {
+    auto task = [this, &result] {
+      result->clear();
+      for (auto& id : set_) {
+        result->push_back(id);
+      }
+      if (VLOG_IS_ON(3)) {
+        std::ostringstream sstream;
+        sstream << "[";
+        for (auto& id : *result) {
+          sstream << id << ", ";
+        }
+        sstream << "]";
+        VLOG(3) << "result ids size: " << result->size() << " "
+                << sstream.str();
+      }
+      set_.clear();
+    };
+    return pool_->enqueue(std::move(task));
+  }
+
+ private:
+  std::unordered_set<int64_t> set_;
+  std::unique_ptr<::ThreadPool> pool_{nullptr};
+};
+
+class AsyncSparseParamUpdateRecorder {
+  using TrainerToRows = std::vector<std::unique_ptr<ConcurrentSet>>;
+
+ public:
+  AsyncSparseParamUpdateRecorder(
+      int trainer_num,
+      const std::unordered_map<std::string, std::string>& grad_to_param)
+      : trainer_num_(trainer_num), grad_to_param_(grad_to_param) {
+    if (VLOG_IS_ON(3)) {
+      std::ostringstream sstream;
+      sstream << "[";
+      for (auto& item : grad_to_param) {
+        sstream << item.first << ":" << item.second << ", ";
+      }
+      sstream << "]";
+      VLOG(3) << "trainer_num: " << trainer_num
+              << " grad_to_param_: " << sstream.str();
+    }
+    for (auto& iter : grad_to_param) {
+      param_to_grad_[iter.second] = iter.first;
+      auto& param_name = iter.second;
+      param_to_updated_rows_[param_name] = TrainerToRows();
+      auto& trainer_to_rows = param_to_updated_rows_[param_name];
+      for (auto i = 0; i < trainer_num; ++i) {
+        trainer_to_rows.emplace_back(new ConcurrentSet());
+      }
+    }
+  }
+
+  ~AsyncSparseParamUpdateRecorder() = default;
+
+  void Update(const std::string& grad_name,
+              const std::vector<int64_t>& update_rows) {
+    VLOG(3) << "update grad: " << grad_name
+            << " row size: " << update_rows.size();
+    auto& param_name = grad_to_param_.at(grad_name);
+    auto& trainer_to_rows = param_to_updated_rows_.at(param_name);
+
+    std::vector<std::future<void>> fs;
+    for (auto& set : trainer_to_rows) {
+      fs.push_back(set->Update(update_rows));
+    }
+    for (auto& f : fs) {
+      f.wait();
+    }
+  }
+
+  void GetAndClear(const std::string& param_name, int trainer_id,
+                   std::vector<int64_t>* result) {
+    VLOG(3) << "GetAndClear param: " << param_name
+            << " for trainer: " << trainer_id;
+    PADDLE_ENFORCE_LT(trainer_id, trainer_num_);
+    param_to_updated_rows_.at(param_name)[trainer_id]
+        ->GetAndClear(result)
+        .wait();
+  }
+
+  bool HasParam(const std::string& param_name) {
+    return param_to_grad_.find(param_name) != param_to_grad_.end();
+  }
+
+  bool HasGrad(const std::string& grad_name) {
+    return grad_to_param_.find(grad_name) != grad_to_param_.end();
+  }
+
+ private:
+  const int trainer_num_;
+  std::unordered_map<std::string, std::string> grad_to_param_;
+  std::unordered_map<std::string, std::string> param_to_grad_;
+  std::unordered_map<std::string, TrainerToRows> param_to_updated_rows_;
+
+  // init recorder
+ public:
+  static void Init(
+      int trainer_num,
+      const std::unordered_map<std::string, std::string>& grad_to_param) {
+    InitImpl(trainer_num, grad_to_param);
+  }
+
+  static AsyncSparseParamUpdateRecorder* GetInstance() {
+    return recorder_.get();
+  }
+
+ private:
+  // Init is called by GetInstance.
+  static void InitImpl(
+      int trainer_num,
+      const std::unordered_map<std::string, std::string>& grad_to_param) {
+    if (recorder_ == nullptr) {
+      recorder_.reset(
+          new AsyncSparseParamUpdateRecorder(trainer_num, grad_to_param));
+    }
+  }
+
+  static std::once_flag init_flag_;
+  static std::unique_ptr<AsyncSparseParamUpdateRecorder> recorder_;
+};
+
+}  // namespace distributed
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/fluid/operators/distributed/async_sparse_param_update_recorder_test.cc b/paddle/fluid/operators/distributed/async_sparse_param_update_recorder_test.cc
new file mode 100644
index 0000000000000000000000000000000000000000..67e8fd8a0edc4510d0abe885c821e75b528254f8
--- /dev/null
+++ b/paddle/fluid/operators/distributed/async_sparse_param_update_recorder_test.cc
@@ -0,0 +1,99 @@
+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/operators/distributed/async_sparse_param_update_recorder.h"
+
+#include <algorithm>
+
+#include "gtest/gtest.h"
+
+namespace paddle {
+namespace operators {
+namespace distributed {
+
+TEST(ConcurrentSet, All) {
+  ConcurrentSet concurrent_set;
+  std::vector<int64_t> in1 = {1, 2, 3, 4};
+  std::vector<int64_t> in2 = {2, 3, 5, 6};
+
+  std::vector<std::future<void>> futures;
+  futures.push_back(concurrent_set.Update(in1));
+  futures.push_back(concurrent_set.Update(in2));
+
+  for (auto &f : futures) {
+    f.wait();
+  }
+
+  std::unordered_set<int64_t> in;
+  std::copy(in1.begin(), in1.end(), std::inserter(in, in.begin()));
+  std::copy(in2.begin(), in2.end(), std::inserter(in, in.begin()));
+
+  std::vector<int64_t> ret;
+  concurrent_set.GetAndClear(&ret).wait();
+
+  std::unordered_set<int64_t> out;
+  std::copy(ret.begin(), ret.end(), std::inserter(out, out.begin()));
+
+  EXPECT_EQ(in, out);
+
+  concurrent_set.GetAndClear(&ret).wait();
+  EXPECT_EQ(ret.size(), 0);
+}
+
+TEST(AsyncSparseParamUpdateRecorder, All) {
+  std::unordered_map<std::string, std::string> grad_to_param;
+  grad_to_param["grad1"] = "param1";
+  grad_to_param["grad2"] = "param2";
+
+  int trainer_num = 10;
+
+  AsyncSparseParamUpdateRecorder recorder(trainer_num, grad_to_param);
+  std::vector<int64_t> in1 = {1, 2, 3, 4};
+  std::vector<int64_t> in2 = {2, 3, 5, 6};
+
+  std::unordered_set<int64_t> in;
+  std::copy(in1.begin(), in1.end(), std::inserter(in, in.begin()));
+  std::copy(in2.begin(), in2.end(), std::inserter(in, in.begin()));
+
+  recorder.Update("grad1", in1);
+  recorder.Update("grad1", in2);
+
+  EXPECT_TRUE(recorder.HasParam("param1"));
+  EXPECT_TRUE(recorder.HasParam("param2"));
+  EXPECT_FALSE(recorder.HasParam("param3"));
+
+  EXPECT_TRUE(recorder.HasGrad("grad1"));
+  EXPECT_TRUE(recorder.HasGrad("grad2"));
+  EXPECT_FALSE(recorder.HasGrad("grad3"));
+
+  std::vector<int64_t> ret;
+  EXPECT_ANY_THROW(recorder.GetAndClear("param1", trainer_num, &ret));
+
+  for (int i = 0; i < trainer_num; ++i) {
+    std::vector<int64_t> ret;
+    std::unordered_set<int64_t> out;
+
+    recorder.GetAndClear("param1", i, &ret);
+    std::copy(ret.begin(), ret.end(), std::inserter(out, out.begin()));
+
+    EXPECT_EQ(in, out);
+
+    recorder.GetAndClear("param1", i, &ret);
+    EXPECT_EQ(ret.size(), 0);
+  }
+}
+
+}  // namespace distributed
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/fluid/operators/distributed/brpc/brpc_client.cc b/paddle/fluid/operators/distributed/brpc/brpc_client.cc
index a1a3443348129b5cdf057592fced8fdff238ac09..4c22ad8eb4d4b2e23d8a6720e726eb9e2998314e 100644
--- a/paddle/fluid/operators/distributed/brpc/brpc_client.cc
+++ b/paddle/fluid/operators/distributed/brpc/brpc_client.cc
@@ -234,6 +234,7 @@ VarHandlePtr BRPCClient::AsyncGetVar(const std::string& ep,
                                      const framework::Scope& scope,
                                      const std::string& var_name,
                                      const std::string& out_var_name,
+                                     const std::string& table_name,
                                      int64_t time_out) {
   return _AsyncGetVar(ep, ctx, scope, var_name, out_var_name, kGetRPC,
                       time_out);
diff --git a/paddle/fluid/operators/distributed/brpc/brpc_client.h b/paddle/fluid/operators/distributed/brpc/brpc_client.h
index 501a593b11d35c160348e42ee47216a85647aac4..51864dfdca53eb4b1d9045188a6347781130e785 100644
--- a/paddle/fluid/operators/distributed/brpc/brpc_client.h
+++ b/paddle/fluid/operators/distributed/brpc/brpc_client.h
@@ -21,8 +21,10 @@ limitations under the License. */
 #include <functional>
 #include <iostream>
 #include <map>
+#include <memory>
 #include <mutex>  // NOLINT
 #include <string>
+#include <unordered_map>
 #include <vector>
 
 #include "brpc/channel.h"
@@ -66,6 +68,7 @@ class BRPCClient : public RPCClient {
                            const framework::Scope& scope,
                            const std::string& var_name,
                            const std::string& out_var_name,
+                           const std::string& table_name = "",
                            int64_t time_out = FLAGS_rpc_deadline) override;
 
   VarHandlePtr AsyncGetMonomerBarrier(
@@ -107,13 +110,11 @@ class BRPCClient : public RPCClient {
   void SendComplete() override;
 
  private:
-  VarHandlePtr _AsyncGetVar(const std::string& ep,
-                            const platform::DeviceContext& ctx,
-                            const framework::Scope& scope,
-                            const std::string& var_name,
-                            const std::string& out_var_name,
-                            const std::string& method_name,
-                            int64_t time_out = FLAGS_rpc_deadline);
+  VarHandlePtr _AsyncGetVar(
+      const std::string& ep, const platform::DeviceContext& ctx,
+      const framework::Scope& scope, const std::string& var_name,
+      const std::string& out_var_name, const std::string& method_name,
+      const std::string& table_name, int64_t time_out = FLAGS_rpc_deadline);
 
   void Proceed();
   ChannelQueuePtr GetChannel(const std::string& ep);
diff --git a/paddle/fluid/operators/distributed/communicator.cc b/paddle/fluid/operators/distributed/communicator.cc
index eba18c67771fa26eed855b0f19591e06101f424d..b528bcdd32b11d686f44596d9a1bb663b21691f4 100644
--- a/paddle/fluid/operators/distributed/communicator.cc
+++ b/paddle/fluid/operators/distributed/communicator.cc
@@ -32,6 +32,9 @@ DEFINE_int32(communicator_send_queue_size, 20,
 DEFINE_int32(communicator_max_send_grad_num_before_recv, 20,
              "max grad num to send before recv parameters");
 DEFINE_int32(communicator_thread_pool_size, 5, "thread num to do send or recv");
+DEFINE_int32(communicator_send_wait_times, 5,
+             "times that send thread will wait if merge num does not reach "
+             "max_merge_var_num");
 DEFINE_int32(communicator_max_merge_var_num, 20,
              "max var num to merge and send");
 DEFINE_bool(communicator_fake_rpc, false,
@@ -65,6 +68,8 @@ Communicator::Communicator(const RpcCtxMap &send_varname_to_ctx,
           << FLAGS_communicator_max_send_grad_num_before_recv;
   VLOG(0) << "communicator_thread_pool_size: "
           << FLAGS_communicator_thread_pool_size;
+  VLOG(0) << "communicator_send_wait_times: "
+          << FLAGS_communicator_send_wait_times;
   VLOG(0) << "communicator_max_merge_var_num: "
           << FLAGS_communicator_max_merge_var_num;
   VLOG(0) << "communicator_fake_rpc: " << FLAGS_communicator_fake_rpc;
@@ -101,20 +106,32 @@ void Communicator::SendThread() {
           VLOG(3) << var_name << " merge and send";
           std::vector<std::shared_ptr<Variable>> vars;
           size_t merged_var_num = 0;
-          while (var_queue->Size() > 0 &&
-                 merged_var_num < FLAGS_communicator_max_merge_var_num) {
-            vars.push_back(var_queue->Pop());
-            // only count the send number of the first var
-            if (var_name == send_varname_to_queue_.begin()->first) {
-              grad_num_.fetch_add(1, std::memory_order_relaxed);
+          size_t wait_times = 0;
+          while (merged_var_num < FLAGS_communicator_max_merge_var_num) {
+            if (var_queue->Size() == 0) {
+              VLOG(3) << "wait_times -> " << wait_times;
+              if (wait_times >= FLAGS_communicator_send_wait_times) {
+                break;
+              }
+              std::this_thread::sleep_for(std::chrono::milliseconds(10));
+              wait_times++;
+              continue;
+            } else {
+              wait_times = 0;
+
+              vars.push_back(var_queue->Pop());
+              // only count the send number of the first var
+              if (var_name == send_varname_to_queue_.begin()->first) {
+                grad_num_.fetch_add(1, std::memory_order_relaxed);
+              }
+              merged_var_num++;
             }
-            merged_var_num++;
           }
           auto before_merge = GetCurrentUS();
           MergeVars(var_name, vars, send_scope_.get());
           auto after_merge = GetCurrentUS();
-          VLOG(3) << "merge " << var_name << " use time "
-                  << after_merge - before_merge;
+          VLOG(3) << "merge " << merged_var_num << " " << var_name
+                  << " use time " << after_merge - before_merge;
           auto send_functor = distributed::ParameterSend<float>();
           auto &ctx = send_varname_to_ctx_.at(var_name);
           if (!FLAGS_communicator_fake_rpc) {
diff --git a/paddle/fluid/operators/distributed/communicator.h b/paddle/fluid/operators/distributed/communicator.h
index 41155bfc31bb31520fdcf5bd50b203f2e1f2c516..37c39eb15112f745f6a25e95ce65d431d825182e 100644
--- a/paddle/fluid/operators/distributed/communicator.h
+++ b/paddle/fluid/operators/distributed/communicator.h
@@ -109,7 +109,7 @@ inline void MergeVars(const std::string& var_name,
   auto* out_var = scope->Var(var_name);
   if (var0->IsType<framework::LoDTensor>()) {
     auto dims = var0->Get<framework::LoDTensor>().dims();
-    VLOG(3) << "merge " << var_name << " LoDTensor " << dims;
+    VLOG(3) << "merge " << var_name << " LoDTensor dims " << dims;
 
     // init output tensor
     auto* out_t = out_var->GetMutable<framework::LoDTensor>();
diff --git a/paddle/fluid/operators/distributed/grpc/grpc_client.cc b/paddle/fluid/operators/distributed/grpc/grpc_client.cc
index 61e94dae3c7a107e10fa5e5518651014cec078bc..8504110c6e9dbfe22b78063999ed4a9e36850e2c 100644
--- a/paddle/fluid/operators/distributed/grpc/grpc_client.cc
+++ b/paddle/fluid/operators/distributed/grpc/grpc_client.cc
@@ -128,9 +128,11 @@ VarHandlePtr GRPCClient::AsyncGetVar(const std::string& ep,
                                      const framework::Scope& scope,
                                      const std::string& var_name,
                                      const std::string& out_varname,
+                                     const std::string& table_name,
                                      int64_t time_out) {
   return _AsyncGetVar(ep, ctx, scope, kGetRPC, var_name, out_varname,
-                      "/sendrecv.SendRecvService/GetVariable", time_out);
+                      "/sendrecv.SendRecvService/GetVariable", table_name,
+                      time_out);
 }
 
 VarHandlePtr GRPCClient::AsyncGetVarNoBarrier(
@@ -142,7 +144,7 @@ VarHandlePtr GRPCClient::AsyncGetVarNoBarrier(
 
   return _AsyncGetVar(
       ep, ctx, scope, kGetNoBarrierRPC, var_name_no_barrier, out_varname,
-      "/sendrecv.SendRecvService/GetVariableNoBarrier", time_out);
+      "/sendrecv.SendRecvService/GetVariableNoBarrier", "", time_out);
 }
 
 VarHandlePtr GRPCClient::AsyncGetMonomerVariable(
@@ -150,18 +152,21 @@ VarHandlePtr GRPCClient::AsyncGetMonomerVariable(
     const framework::Scope& scope, const std::string& var_name,
     int64_t time_out) {
   return _AsyncGetVar(ep, ctx, scope, kGetMonomerRPC, var_name, var_name,
-                      "/sendrecv.SendRecvService/GetMonomerVariable", time_out);
+                      "/sendrecv.SendRecvService/GetMonomerVariable", "",
+                      time_out);
 }
 
 VarHandlePtr GRPCClient::_AsyncGetVar(
     const std::string& ep, const platform::DeviceContext& ctx,
     const framework::Scope& scope, const std::string& method,
     const std::string& var_name, const std::string& out_varname,
-    const std::string& rpc_path, int64_t time_out) {
+    const std::string& rpc_path, const std::string& table_name,
+    int64_t time_out) {
   const platform::DeviceContext* p_ctx = &ctx;
   const std::string ep_val = ep;
   const std::string var_name_val = var_name;
   const std::string out_varname_val = out_varname;
+  const std::string table_name_val = table_name;
   const framework::Scope* p_scope = &scope;
   const auto ch = GetChannel(ep_val);
   GetProcessor* s = new GetProcessor(ch);
@@ -169,32 +174,33 @@ VarHandlePtr GRPCClient::_AsyncGetVar(
   VarHandlePtr h(new VarHandle(ep, method, out_varname_val, p_ctx, p_scope));
   s->Prepare(h, time_out);
 
-  framework::AsyncIO(
-      [var_name_val, out_varname_val, s, method, p_ctx, h, rpc_path, this] {
-        // prepare input
-        sendrecv::VariableMessage req;
-        req.set_varname(var_name_val);
-        req.set_out_varname(out_varname_val);
-        req.set_trainer_id(trainer_id_);
-        ::grpc::ByteBuffer buf;
-        RequestToByteBuffer<sendrecv::VariableMessage>(req, &buf);
+  framework::AsyncIO([var_name_val, out_varname_val, table_name_val, s, method,
+                      p_ctx, h, rpc_path, this] {
+    // prepare input
+    sendrecv::VariableMessage req;
+    req.set_varname(var_name_val);
+    req.set_out_varname(out_varname_val);
+    req.set_trainer_id(trainer_id_);
+    req.set_table_name(table_name_val);
+    ::grpc::ByteBuffer buf;
+    RequestToByteBuffer<sendrecv::VariableMessage>(req, &buf);
 
-        VLOG(3) << s->GetVarHandlePtr()->String() << " begin";
+    VLOG(3) << s->GetVarHandlePtr()->String() << " begin";
 
-        // stub context
-        s->response_call_back_ = ProcGetResponse;
+    // stub context
+    s->response_call_back_ = ProcGetResponse;
 
-        platform::RecordRPCEvent record_event(method);
+    platform::RecordRPCEvent record_event(method);
 
-        auto call =
-            s->stub_g_.PrepareUnaryCall(s->context_.get(), rpc_path, buf, &cq_);
-        call->StartCall();
-        call->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
+    auto call =
+        s->stub_g_.PrepareUnaryCall(s->context_.get(), rpc_path, buf, &cq_);
+    call->StartCall();
+    call->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
 
-        if (UNLIKELY(platform::IsProfileEnabled())) {
-          h->Wait();
-        }
-      });
+    if (UNLIKELY(platform::IsProfileEnabled())) {
+      h->Wait();
+    }
+  });
 
   req_count_++;
 
diff --git a/paddle/fluid/operators/distributed/grpc/grpc_client.h b/paddle/fluid/operators/distributed/grpc/grpc_client.h
index ce0d2152aa27c62b6e12881aaf2ae458597e67e6..ad2f04a6d1dda34e35b67b21dce8ac612ff697a0 100644
--- a/paddle/fluid/operators/distributed/grpc/grpc_client.h
+++ b/paddle/fluid/operators/distributed/grpc/grpc_client.h
@@ -23,9 +23,11 @@ limitations under the License. */
 #include <functional>
 #include <iostream>
 #include <map>
+#include <memory>
 #include <mutex>  // NOLINT
 #include <string>
 #include <thread>  // NOLINT
+#include <unordered_map>
 #include <vector>
 
 #include "grpc++/channel.h"
@@ -187,6 +189,7 @@ class GRPCClient : public RPCClient {
                            const framework::Scope& scope,
                            const std::string& var_name,
                            const std::string& out_varname,
+                           const std::string& table_name = "",
                            int64_t time_out = FLAGS_rpc_deadline) override;
 
   VarHandlePtr AsyncGetVarNoBarrier(
@@ -239,7 +242,8 @@ class GRPCClient : public RPCClient {
       const std::string& ep, const platform::DeviceContext& ctx,
       const framework::Scope& scope, const std::string& method,
       const std::string& var_name, const std::string& out_varname,
-      const std::string& rpc_path, int64_t time_out = FLAGS_rpc_deadline);
+      const std::string& rpc_path, const std::string& table_name = "",
+      int64_t time_out = FLAGS_rpc_deadline);
 
  private:
   grpc::CompletionQueue cq_;
diff --git a/paddle/fluid/operators/distributed/grpc/grpc_server.cc b/paddle/fluid/operators/distributed/grpc/grpc_server.cc
index 0eb313f75dfa64f8722faa365128f3111f72bd0b..75526bed0f0eadada65279ec05757da7a469f984 100644
--- a/paddle/fluid/operators/distributed/grpc/grpc_server.cc
+++ b/paddle/fluid/operators/distributed/grpc/grpc_server.cc
@@ -137,6 +137,7 @@ class RequestGet final : public RequestBase {
     // proc request.
     std::string varname = request_.varname();
     std::string out_varname = request_.out_varname();
+    std::string table_name = request_.table_name();
     int trainer_id = request_.trainer_id();
 
     VLOG(4) << "RequestGet " << out_varname << " from " << varname;
@@ -145,19 +146,23 @@ class RequestGet final : public RequestBase {
     framework::Variable* invar = nullptr;
     framework::Variable* outvar = nullptr;
 
-    request_handler_->Handle(varname, scope, invar, &outvar, trainer_id,
-                             out_varname);
+    tmp_scope_ = std::move(scope->NewTmpScope());
+    request_handler_->Handle(varname, tmp_scope_.get(), invar, &outvar,
+                             trainer_id, out_varname, table_name);
 
+    VLOG(1) << "before SerializeToByteBuffer";
     if (outvar) {
       SerializeToByteBuffer(out_varname, outvar, *request_handler_->dev_ctx(),
                             &reply_);
     }
+    VLOG(1) << "after SerializeToByteBuffer";
     Finish(reply_, &responder_);
   }
 
  protected:
   sendrecv::VariableMessage request_;
   ::grpc::ByteBuffer reply_;
+  std::unique_ptr<framework::Scope> tmp_scope_;
   ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_;
 };
 
diff --git a/paddle/fluid/operators/distributed/parameter_recv.cc b/paddle/fluid/operators/distributed/parameter_recv.cc
index e7d4c262aa9fad10a23adc61b94ba0c38577c0e8..da73167ae603fb8c8ba9deabe118269891d1f52a 100644
--- a/paddle/fluid/operators/distributed/parameter_recv.cc
+++ b/paddle/fluid/operators/distributed/parameter_recv.cc
@@ -42,27 +42,23 @@ using DDim = framework::DDim;
 template <typename T>
 void ParameterRecv<T>::operator()(const RpcContext &rpc_ctx,
                                   const framework::Scope &scope) {
-  VLOG(3) << "ParameterRecv in";
+  VLOG(3) << "ParameterRecv in " << rpc_ctx.var_name;
   std::unique_ptr<framework::Scope> local_scope = scope.NewTmpScope();
 
   platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
   auto &cpu_ctx = *pool.Get(platform::CPUPlace());
 
   distributed::RPCClient *rpc_client =
-      distributed::RPCClient::GetInstance<RPCCLIENT_T>(0);
+      distributed::RPCClient::GetInstance<RPCCLIENT_T>(rpc_ctx.trainer_id);
 
   auto *recv_var = scope.FindVar(rpc_ctx.var_name);
 
-  std::vector<framework::Tensor *> recved_tensors;
-
   // recv all vars to local scope
   if (recv_var->IsType<framework::LoDTensor>()) {
     std::vector<distributed::VarHandlePtr> rets;
     for (size_t i = 0; i < rpc_ctx.splited_var_names.size(); i++) {
       auto &recv_var_name = rpc_ctx.splited_var_names[i];
-      framework::Tensor *t =
-          local_scope->Var(recv_var_name)->GetMutable<framework::LoDTensor>();
-      recved_tensors.push_back(t);
+      local_scope->Var(recv_var_name);
       VLOG(3) << "recv " << recv_var_name << " from " << rpc_ctx.epmap[i];
       rets.push_back(rpc_client->AsyncGetVar(rpc_ctx.epmap[i], cpu_ctx,
                                              *local_scope.get(), recv_var_name,
@@ -78,23 +74,61 @@ void ParameterRecv<T>::operator()(const RpcContext &rpc_ctx,
   // concat recved tensor into one var
   {
     size_t output_offset = 0;
+    size_t row_offset = 0;
     framework::Tensor *recv_tensor =
         recv_var->GetMutable<framework::LoDTensor>();
     auto dev_ctx = paddle::platform::CPUDeviceContext();
     int64_t recv_numel = 0;
-    for (auto *in : recved_tensors) {
-      recv_numel += in->numel();
-      auto in_stride = framework::stride_numel(in->dims());
-      auto out_stride = framework::stride_numel(recv_tensor->dims());
-      StridedNumelCopyWithAxis<T>(
-          dev_ctx, 0, recv_tensor->data<T>() + output_offset, out_stride,
-          in->data<T>(), in_stride, in_stride[0]);
-      output_offset += in_stride[0];
+    for (auto &recv_var_name : rpc_ctx.splited_var_names) {
+      auto *recv_var = local_scope->FindVar(recv_var_name);
+      if (recv_var->IsType<framework::LoDTensor>()) {
+        auto &in = recv_var->Get<framework::LoDTensor>();
+        recv_numel += in.numel();
+        auto in_stride = framework::stride_numel(in.dims());
+        auto out_stride = framework::stride_numel(recv_tensor->dims());
+        StridedNumelCopyWithAxis<T>(
+            dev_ctx, 0, recv_tensor->data<T>() + output_offset, out_stride,
+            in.data<T>(), in_stride, in_stride[0]);
+        output_offset += in_stride[0];
+      } else if (recv_var->IsType<framework::SelectedRows>()) {
+        auto &recv_slr = recv_var->Get<framework::SelectedRows>();
+        auto &recv_dims = recv_tensor->dims();
+        int64_t width = recv_dims[1];
+        recv_numel += recv_slr.height() * width;
+        PADDLE_ENFORCE_EQ(recv_slr.value().dims()[1], width);
+        PADDLE_ENFORCE_EQ(recv_slr.value().dims()[0], recv_slr.rows().size());
+        VLOG(3) << "recv slr " << recv_var_name << " dims "
+                << recv_slr.value().dims();
+        if (VLOG_IS_ON(3)) {
+          std::ostringstream sstream;
+          sstream << "[";
+          for (auto &row_id : recv_slr.rows()) {
+            sstream << row_id << ", ";
+          }
+          sstream << "]";
+          VLOG(3) << "recv_slr size: " << recv_slr.rows().size() << " "
+                  << sstream.str();
+        }
+
+        for (auto i = 0; i < recv_slr.rows().size(); ++i) {
+          auto row_id = recv_slr.rows()[i] + row_offset;
+          PADDLE_ENFORCE_LT(row_id, recv_dims[0]);
+          memcpy(recv_tensor->data<T>() + row_id * width,
+                 recv_slr.value().data<T>() + i * width, sizeof(T) * width);
+        }
+        row_offset += recv_slr.height();
+      } else {
+        PADDLE_THROW("unsupported recieved var type");
+      }
+    }
+    auto numel = recv_tensor->numel();
+    if (recv_numel != numel) {
+      LOG(FATAL) << "recv_numel: " << recv_numel << " acture numel: " << numel;
     }
-    PADDLE_ENFORCE_EQ(recv_numel, recv_tensor->numel());
+    PADDLE_ENFORCE_EQ(recv_numel, numel);
   }
 
-  VLOG(3) << "ParameterRecv out";
+  VLOG(3) << "ParameterRecv out " << rpc_ctx.var_name;
 }
 
 template struct ParameterRecv<float>;
diff --git a/paddle/fluid/operators/distributed/parameter_send.cc b/paddle/fluid/operators/distributed/parameter_send.cc
index 9ce424445229cde0a7e775c95f4af8839f4d4d68..dfabad567af590b65b9e777824d476fce2b17238 100644
--- a/paddle/fluid/operators/distributed/parameter_send.cc
+++ b/paddle/fluid/operators/distributed/parameter_send.cc
@@ -47,7 +47,7 @@ void ParameterSend<T>::operator()(const RpcContext &rpc_ctx,
   auto &cpu_ctx = *pool.Get(platform::CPUPlace());
 
   distributed::RPCClient *rpc_client =
-      distributed::RPCClient::GetInstance<RPCCLIENT_T>(0);
+      distributed::RPCClient::GetInstance<RPCCLIENT_T>(rpc_ctx.trainer_id);
 
   auto *send_var = scope.FindVar(rpc_ctx.var_name);
   size_t out_num = rpc_ctx.splited_var_names.size();
diff --git a/paddle/fluid/operators/distributed/request_handler.h b/paddle/fluid/operators/distributed/request_handler.h
index 991158ac72007efc1233f852caed4f90f35fe1cd..de8f30184611aeb961e2ab69b05779c56371b976 100644
--- a/paddle/fluid/operators/distributed/request_handler.h
+++ b/paddle/fluid/operators/distributed/request_handler.h
@@ -18,7 +18,9 @@
 #include <condition_variable>  // NOLINT
 
 #include <functional>
+#include <memory>
 #include <string>
+#include <unordered_map>
 #include <utility>
 #include <vector>
 
@@ -180,6 +182,10 @@ class RequestHandler {
     grad_to_prepared_ctx_ = g;
   }
 
+  void SetSparseGradToParam(std::unordered_map<std::string, std::string>* g) {
+    sparse_grad_to_param_ = g;
+  }
+
   void SetRPCServer(RPCServer* rpc_server) { rpc_server_ = rpc_server; }
 
   // Get attributes.
@@ -228,6 +234,7 @@ class RequestHandler {
   std::unordered_map<std::string,
                      std::shared_ptr<framework::ExecutorPrepareContext>>*
       grad_to_prepared_ctx_;
+  std::unordered_map<std::string, std::string>* sparse_grad_to_param_;
 
   RPCServer* rpc_server_;
 };
diff --git a/paddle/fluid/operators/distributed/request_handler_impl.cc b/paddle/fluid/operators/distributed/request_handler_impl.cc
index e289ec929dbd6643a2518b92c1a25b7d63e790a9..a41536368abc925531d1a54615546a100482a7eb 100644
--- a/paddle/fluid/operators/distributed/request_handler_impl.cc
+++ b/paddle/fluid/operators/distributed/request_handler_impl.cc
@@ -22,6 +22,7 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/framework/variable_helper.h"
+#include "paddle/fluid/operators/distributed/async_sparse_param_update_recorder.h"
 #include "paddle/fluid/operators/distributed/rpc_server.h"
 #include "paddle/fluid/string/piece.h"
 #include "paddle/fluid/string/printf.h"
@@ -59,6 +60,12 @@ bool RequestSendHandler::Handle(const std::string& varname,
             "async mode should not recv BATCH_BARRIER_MESSAGE or "
             "COMPLETE_MESSAGE");
       }
+      if (AsyncSparseParamUpdateRecorder::GetInstance()->HasGrad(varname)) {
+        auto& grad_slr =
+            scope->FindVar(varname)->Get<framework::SelectedRows>();
+        AsyncSparseParamUpdateRecorder::GetInstance()->Update(varname,
+                                                              grad_slr.rows());
+      }
       executor_->RunPreparedContext((*grad_to_prepared_ctx_)[varname].get(),
                                     scope);
       return true;
@@ -82,8 +89,9 @@ bool RequestGetHandler::Handle(const std::string& varname,
                                const int trainer_id,
                                const std::string& out_var_name,
                                const std::string& table_name) {
-  VLOG(4) << "RequestGetHandler:" << varname
-          << " out_var_name: " << out_var_name;
+  VLOG(3) << "RequestGetHandler:" << varname
+          << " out_var_name: " << out_var_name << " trainer_id: " << trainer_id
+          << " table_name: " << table_name;
 
   if (sync_mode_) {
     if (varname == FETCH_BARRIER_MESSAGE) {
@@ -108,7 +116,42 @@ bool RequestGetHandler::Handle(const std::string& varname,
         VLOG(3) << "copying " << varname << " to " << param_bak_name;
         framework::TensorCopy(t_orig, dev_ctx_->GetPlace(), t);
       }
-      *outvar = scope_->FindVar(varname);
+      if (AsyncSparseParamUpdateRecorder::GetInstance()->HasParam(varname) &&
+          !table_name.empty()) {
+        std::vector<int64_t> updated_rows;
+        AsyncSparseParamUpdateRecorder::GetInstance()->GetAndClear(
+            varname, trainer_id, &updated_rows);
+        if (VLOG_IS_ON(3)) {
+          std::ostringstream sstream;
+          sstream << "[";
+          for (auto& row_id : updated_rows) {
+            sstream << row_id << ", ";
+          }
+          sstream << "]";
+          VLOG(3) << "updated_rows size: " << updated_rows.size() << " "
+                  << sstream.str();
+        }
+        auto& origin_tensor =
+            scope_->FindVar(varname)->Get<framework::LoDTensor>();
+        auto* origin_tensor_data = origin_tensor.data<float>();
+        auto& dims = origin_tensor.dims();
+        *outvar = scope->Var();
+        auto* out_slr = (*outvar)->GetMutable<framework::SelectedRows>();
+        out_slr->set_rows(updated_rows);
+        out_slr->set_height(dims[0]);
+        auto out_dims = framework::make_ddim(
+            {static_cast<int64_t>(updated_rows.size()), dims[1]});
+        auto* data = out_slr->mutable_value()->mutable_data<float>(
+            out_dims, origin_tensor.place());
+        auto width = dims[1];
+        for (auto i = 0; i < updated_rows.size(); ++i) {
+          PADDLE_ENFORCE_LT(updated_rows[i], dims[0]);
+          memcpy(data + i * width, origin_tensor_data + updated_rows[i] * width,
+                 sizeof(float) * width);
+        }
+      } else {
+        *outvar = scope_->FindVar(varname);
+      }
     }
   }
   return true;
diff --git a/paddle/fluid/operators/distributed/rpc_client.h b/paddle/fluid/operators/distributed/rpc_client.h
index ea54e0c2951253fc009672f4cd2e5233ed56944e..d4be2c28fdbaa4beef62402155de5b677ed67e9b 100644
--- a/paddle/fluid/operators/distributed/rpc_client.h
+++ b/paddle/fluid/operators/distributed/rpc_client.h
@@ -15,6 +15,7 @@
 #pragma once
 
 #include <condition_variable>  // NOLINT
+#include <memory>
 #include <string>
 #include "gflags/gflags.h"
 
@@ -44,6 +45,7 @@ class RPCClient {
                                    const framework::Scope& scope,
                                    const std::string& var_name,
                                    const std::string& out_varname,
+                                   const std::string& table_name = "",
                                    int64_t time_out = FLAGS_rpc_deadline) = 0;
 
   virtual VarHandlePtr AsyncGetVarNoBarrier(
@@ -96,6 +98,7 @@ class RPCClient {
   // Init is called by GetInstance.
   template <typename T>
   static void Init(int trainer_id) {
+    VLOG(0) << "init rpc client with trainer_id " << trainer_id;
     trainer_id_ = trainer_id;
     if (rpc_client_.get() == nullptr) {
       rpc_client_.reset(new T());
diff --git a/paddle/fluid/operators/distributed/rpc_common.h b/paddle/fluid/operators/distributed/rpc_common.h
index 3de89c2ae89d29edc317ca123882d1c55038b6ca..eb127bf4ad5a5c9a28210e2fbcdb69b07543f4b9 100644
--- a/paddle/fluid/operators/distributed/rpc_common.h
+++ b/paddle/fluid/operators/distributed/rpc_common.h
@@ -27,23 +27,26 @@ struct RpcContext {
 
   RpcContext(const std::string &name, const std::vector<std::string> &names,
              const std::vector<std::string> &emap,
-             const std::vector<int64_t> &sections)
+             const std::vector<int64_t> &sections, int id)
       : var_name(name),
         splited_var_names(names),
         epmap(emap),
-        height_sections(sections) {}
+        height_sections(sections),
+        trainer_id(id) {}
 
   RpcContext(const RpcContext &ctx) {
     var_name = ctx.var_name;
     splited_var_names = ctx.splited_var_names;
     epmap = ctx.epmap;
     height_sections = ctx.height_sections;
+    trainer_id = ctx.trainer_id;
   }
 
   std::string var_name;
   std::vector<std::string> splited_var_names;
   std::vector<std::string> epmap;
   std::vector<int64_t> height_sections;
+  int trainer_id;
 };
 
 inline std::ostream &operator<<(std::ostream &os, const RpcContext &rpc_ctx) {
diff --git a/paddle/fluid/operators/distributed_ops/CMakeLists.txt b/paddle/fluid/operators/distributed_ops/CMakeLists.txt
index a1ef1af39ff2ab1456706ebafbd3d7ce1acc0c07..1096f3773c6d44560d370502b1c550d67d40ca64 100644
--- a/paddle/fluid/operators/distributed_ops/CMakeLists.txt
+++ b/paddle/fluid/operators/distributed_ops/CMakeLists.txt
@@ -2,9 +2,9 @@ include(operators)
 
 set(DISTRIBUTE_DEPS "")
 if(WITH_GRPC)
-    set(DISTRIBUTE_DEPS sendrecvop_rpc parameter_send parameter_recv communicator grpc++_unsecure grpc_unsecure gpr cares zlib protobuf node)
+    set(DISTRIBUTE_DEPS sendrecvop_rpc parameter_send parameter_recv communicator async_sparse_param_update_recorder grpc++_unsecure grpc_unsecure gpr cares zlib protobuf node)
 else()
-    set(DISTRIBUTE_DEPS sendrecvop_rpc parameter_send parameter_recv communicator brpc leveldb snappystream snappy protobuf ssl crypto zlib node)
+    set(DISTRIBUTE_DEPS sendrecvop_rpc parameter_send parameter_recv communicator async_sparse_param_update_recorder brpc leveldb snappystream snappy protobuf ssl crypto zlib node)
     if(WITH_BRPC_RDMA)
         find_library(IBVERBS_LIBRARY NAMES ibverbs)
         ADD_LIBRARY(ibverbs SHARED IMPORTED GLOBAL)
diff --git a/paddle/fluid/operators/distributed_ops/listen_and_serv_op.cc b/paddle/fluid/operators/distributed_ops/listen_and_serv_op.cc
index 5b30ed472d51a37a0705d1717395da9e4ff7d743..a672fb2a9141a81383d947dcc961a112aee3f7ac 100644
--- a/paddle/fluid/operators/distributed_ops/listen_and_serv_op.cc
+++ b/paddle/fluid/operators/distributed_ops/listen_and_serv_op.cc
@@ -24,8 +24,10 @@ limitations under the License. */
 #include "paddle/fluid/operators/distributed/distributed.h"
 #include "paddle/fluid/operators/math/math_function.h"
 
+#include "paddle/fluid/operators/distributed/async_sparse_param_update_recorder.h"
 #include "paddle/fluid/operators/distributed/request_handler_impl.h"
 #include "paddle/fluid/operators/distributed_ops/listen_and_serv_op.h"
+
 #include "paddle/fluid/platform/profiler.h"
 
 DEFINE_int32(rpc_send_thread_num, 12, "number of threads for rpc send");
@@ -292,6 +294,8 @@ static void FillRequestCtx(
     std::unordered_map<std::string,
                        std::shared_ptr<framework::ExecutorPrepareContext>>
         *prefetch_ctx,
+    std::unordered_map<std::string, std::string>
+        *sparse_grad_name_to_param_name,
     std::shared_ptr<framework::ExecutorPrepareContext> checkpoint_ctx,
     distributed::RPCServer *rpc_server) {
   h->SetScope(scope);
@@ -299,6 +303,7 @@ static void FillRequestCtx(
   h->SetExecutor(executor);
   h->SetProgram(program);
   h->SetPrefetchPreparedCtx(prefetch_ctx);
+  h->SetSparseGradToParam(sparse_grad_name_to_param_name);
   h->SetRPCServer(rpc_server);
   h->SetCheckpointNotifyPreparedCtx(checkpoint_ctx);
 }
@@ -414,10 +419,24 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
     prefetch_var_name_to_prepared_ctx[prefetch_var_name] = prefetch_prepared[i];
   }
 
-  auto f =
-      std::bind(FillRequestCtx, std::placeholders::_1, &recv_scope, &dev_ctx,
-                &executor, program, &prefetch_var_name_to_prepared_ctx,
-                ckpt_pre_context, rpc_service_.get());
+  // parse attr of kSparseGradToParam  sparse_grad_name -> param_name
+  std::unordered_map<std::string, std::string> sparse_grad_name_to_param_name;
+  auto sparse_grad_name_to_param_name_str =
+      Attr<std::vector<std::string>>(kSparseGradToParam);
+  for (const auto &sparse_grad_name_and_param_name :
+       sparse_grad_name_to_param_name_str) {
+    std::vector<std::string> pieces;
+    split(sparse_grad_name_and_param_name, ':', &pieces);
+    PADDLE_ENFORCE_EQ(pieces.size(), 2);
+    VLOG(3) << "after split, sparse_grad_name = " << pieces[0]
+            << ", param_name = " << pieces[1];
+    sparse_grad_name_to_param_name[pieces[0]] = pieces[1];
+  }
+
+  auto f = std::bind(
+      FillRequestCtx, std::placeholders::_1, &recv_scope, &dev_ctx, &executor,
+      program, &prefetch_var_name_to_prepared_ctx,
+      &sparse_grad_name_to_param_name, ckpt_pre_context, rpc_service_.get());
 
   f(request_send_handler_.get());
   f(request_get_handler_.get());
@@ -445,6 +464,8 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
     RunSyncLoop(&executor, program, &recv_scope, &dev_ctx,
                 prefetch_block_id_list, checkpoint_block_id);
   } else {
+    distributed::AsyncSparseParamUpdateRecorder::Init(
+        fan_in, sparse_grad_name_to_param_name);
     RunAsyncLoop(&executor, program, &recv_scope);
   }
 }
@@ -475,6 +496,10 @@ class ListenAndServOpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<std::vector<std::string>>(kPrefetchVarNameToBlockId,
                                       "prefetch blocks to run on server side.")
         .SetDefault({});
+    AddAttr<std::vector<std::string>>(
+        kSparseGradToParam,
+        "sparse grad name to param name. like: 'emb@Grad:emb'")
+        .SetDefault({});
     AddAttr<int>("Fanin", "How many clients send to this server.")
         .SetDefault(1);
     AddAttr<int>(kCheckpointBlockId,
diff --git a/paddle/fluid/operators/distributed_ops/listen_and_serv_op.h b/paddle/fluid/operators/distributed_ops/listen_and_serv_op.h
index f20442bad7c5bd96173b9d6efc4dceb13feacf5b..1cf2130d7a593077d1145b4f3be379c32557dd53 100644
--- a/paddle/fluid/operators/distributed_ops/listen_and_serv_op.h
+++ b/paddle/fluid/operators/distributed_ops/listen_and_serv_op.h
@@ -16,8 +16,10 @@ limitations under the License. */
 
 #include <stdint.h>
 #include <atomic>
+#include <memory>
 #include <set>
 #include <string>
+#include <unordered_map>
 #include <utility>
 #include <vector>
 
@@ -35,6 +37,7 @@ namespace operators {
 constexpr char kOptimizeBlocks[] = "optimize_blocks";
 constexpr char kPrefetchVarNameToBlockId[] = "prefetch_var_name_to_block_id";
 constexpr char kCheckpointBlockId[] = "checkpint_block_id";
+constexpr char kSparseGradToParam[] = "sparse_grad_to_param";
 
 void RunServer(std::shared_ptr<distributed::RPCServer> service);
 
diff --git a/paddle/fluid/operators/distributed_ops/recv_op.cc b/paddle/fluid/operators/distributed_ops/recv_op.cc
index 3fd0700a077321d931e87b1d94c3637d167c9eff..8e9846b1fc89953526149be3838103526d5c441b 100644
--- a/paddle/fluid/operators/distributed_ops/recv_op.cc
+++ b/paddle/fluid/operators/distributed_ops/recv_op.cc
@@ -50,17 +50,18 @@ class RecvOp : public framework::OperatorBase {
 
     platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
     auto &ctx = *pool.Get(place);
+    auto trainer_id = Attr<int>("trainer_id");
 
     distributed::RPCClient *rpc_client =
-        distributed::RPCClient::GetInstance<RPCCLIENT_T>(
-            Attr<int>("trainer_id"));
+        distributed::RPCClient::GetInstance<RPCCLIENT_T>(trainer_id);
 
     std::vector<std::string> recv_varnames =
         Attr<std::vector<std::string>>("recv_varnames");
 
     if (recv_varnames.size() > 0) {
       auto recv_functor = distributed::ParameterRecv<float>();
-      auto rpc_ctx = distributed::RpcContext(outs[0], recv_varnames, epmap, {});
+      auto rpc_ctx = distributed::RpcContext(outs[0], recv_varnames, epmap, {},
+                                             trainer_id);
       recv_functor(rpc_ctx, scope);
     } else {
       if (with_barrier) {
diff --git a/paddle/fluid/operators/distributed_ops/send_op.cc b/paddle/fluid/operators/distributed_ops/send_op.cc
index b08cd0942f8c89b60d722c931d0cec2063b96578..5731bcc15a07074b3d77873c5cdcbb70dc41aba8 100644
--- a/paddle/fluid/operators/distributed_ops/send_op.cc
+++ b/paddle/fluid/operators/distributed_ops/send_op.cc
@@ -42,6 +42,7 @@ class SendOp : public framework::OperatorBase {
 
     auto epmap = Attr<std::vector<std::string>>("epmap");
     int sync_send = Attr<int>("sync_mode");
+    auto trainer_id = Attr<int>("trainer_id");
 
     auto send_varnames = Attr<std::vector<std::string>>("send_varnames");
     auto height_sections = Attr<std::vector<int64_t>>("sections");
@@ -51,7 +52,7 @@ class SendOp : public framework::OperatorBase {
       if (distributed::Communicator::GetInstance() == nullptr) {
         auto send_functor = distributed::ParameterSend<float>();
         auto rpc_ctx = distributed::RpcContext(ins[0], send_varnames, epmap,
-                                               height_sections);
+                                               height_sections, trainer_id);
         send_functor(rpc_ctx, scope, true);
       } else {
         distributed::Communicator::GetInstance()->Send(ins[0], scope);
@@ -62,8 +63,7 @@ class SendOp : public framework::OperatorBase {
       auto& ctx = *pool.Get(place);
 
       distributed::RPCClient* rpc_client =
-          distributed::RPCClient::GetInstance<RPCCLIENT_T>(
-              Attr<int>("trainer_id"));
+          distributed::RPCClient::GetInstance<RPCCLIENT_T>(trainer_id);
 
       std::vector<distributed::VarHandlePtr> rets;
       for (size_t i = 0; i < ins.size(); i++) {
diff --git a/paddle/fluid/operators/linspace_op.cc b/paddle/fluid/operators/linspace_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..f4aeb062d8dfae31a72b8ebccb3d377276662da6
--- /dev/null
+++ b/paddle/fluid/operators/linspace_op.cc
@@ -0,0 +1,84 @@
+/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/linspace_op.h"
+
+namespace paddle {
+namespace operators {
+
+class LinspaceOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Start"),
+                   "Input(Start) of LinspaceOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Stop"),
+                   "Input(Stop) of LinspaceOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Num"),
+                   "Input(Num) of LinspaceOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(OUt) of LinspaceOp should not be null.");
+
+    auto s_dims = ctx->GetInputDim("Start");
+    PADDLE_ENFORCE((s_dims.size() == 1) && (s_dims[0] == 1),
+                   "The shape of Input(Start) should be [1].");
+
+    auto e_dims = ctx->GetInputDim("Stop");
+    PADDLE_ENFORCE((e_dims.size() == 1) && (e_dims[0] == 1),
+                   "The shape of Input(Stop) should be [1].");
+
+    auto step_dims = ctx->GetInputDim("Num");
+    PADDLE_ENFORCE((step_dims.size() == 1) && (step_dims[0] == 1),
+                   "The shape of Input(Num) should be [1].");
+
+    ctx->SetOutputDim("Out", {-1});
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    framework::LibraryType library_{framework::LibraryType::kPlain};
+    framework::DataLayout layout_ = framework::DataLayout::kAnyLayout;
+    return framework::OpKernelType(
+        ctx.Input<framework::Tensor>("Start")->type(), ctx.device_context(),
+        layout_, library_);
+  }
+};
+
+class LinspaceOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("Start",
+             "First entry in the sequence. It is a tensor of shape [1], should "
+             "be of type float32 or float64.");
+    AddInput("Stop",
+             "Last entry in the sequence. It is a tensor of shape [1], should "
+             "be of type float32 or float64.");
+    AddInput("Num",
+             "Number of entry in the sequence. It is a tensor of shape [1], "
+             "should be of type int32.");
+    AddOutput("Out", "A sequence of numbers.");
+    AddComment(R"DOC(
+    Return fixed number of evenly spaced values within a given interval. First entry is start, and last entry is stop. In the case when Num is 1, only Start is returned. Like linspace function of numpy.
+)DOC");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(linspace, ops::LinspaceOp, ops::LinspaceOpMaker);
+REGISTER_OP_CPU_KERNEL(linspace, ops::CPULinspaceKernel<float>,
+                       ops::CPULinspaceKernel<double>);
diff --git a/paddle/fluid/operators/linspace_op.cu b/paddle/fluid/operators/linspace_op.cu
new file mode 100644
index 0000000000000000000000000000000000000000..90bd17cda0e0d1f78810233537bb502f9115fbd0
--- /dev/null
+++ b/paddle/fluid/operators/linspace_op.cu
@@ -0,0 +1,75 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/linspace_op.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
+
+namespace paddle {
+namespace operators {
+
+#define CUDA_1D_KERNEL_LOOP(i, n)                              \
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \
+       i += blockDim.x * gridDim.x)
+
+template <typename T>
+__global__ void LinspaceKernel(T start, T step, int64_t size, T* out) {
+  CUDA_1D_KERNEL_LOOP(index, size) { out[index] = start + step * index; }
+}
+
+template <typename T>
+__global__ void LinspaceSpecialKernel(T start, T* out) {
+  out[0] = start;
+}
+
+template <typename T>
+class CUDALinspaceKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* start_t = context.Input<framework::Tensor>("Start");
+    auto* stop_t = context.Input<framework::Tensor>("Stop");
+    auto* num_t = context.Input<framework::Tensor>("Num");
+    auto* out = context.Output<framework::Tensor>("Out");
+
+    framework::Tensor n;
+    framework::TensorCopy(*start_t, platform::CPUPlace(), &n);
+    T start = n.data<T>()[0];
+    framework::TensorCopy(*stop_t, platform::CPUPlace(), &n);
+    T stop = n.data<T>()[0];
+    framework::TensorCopy(*num_t, platform::CPUPlace(), &n);
+    int32_t num = n.data<int32_t>()[0];
+
+    PADDLE_ENFORCE(num > 0, "The num of linspace op should be larger than 0.");
+
+    out->Resize(framework::make_ddim({num}));
+    T* out_data = out->mutable_data<T>(context.GetPlace());
+
+    T step = 0;
+    if (num != 1) {
+      step = (stop - start) / (num - 1);
+    }
+
+    auto stream = context.cuda_device_context().stream();
+    int block = 512;
+    int grid = (num + block - 1) / block;
+    LinspaceKernel<T><<<grid, block, 0, stream>>>(start, step, num, out_data);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(linspace, ops::CUDALinspaceKernel<float>,
+                        ops::CUDALinspaceKernel<double>);
diff --git a/paddle/fluid/operators/linspace_op.h b/paddle/fluid/operators/linspace_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..b1fcac73b0ad249aa19859bde770a8554cdb7408
--- /dev/null
+++ b/paddle/fluid/operators/linspace_op.h
@@ -0,0 +1,51 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include <functional>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+class CPULinspaceKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    T start = context.Input<framework::Tensor>("Start")->data<T>()[0];
+    T stop = context.Input<framework::Tensor>("Stop")->data<T>()[0];
+    int32_t num = context.Input<framework::Tensor>("Num")->data<int32_t>()[0];
+    auto* out = context.Output<framework::Tensor>("Out");
+    PADDLE_ENFORCE(num > 0, "The num of linspace op should be larger than 0.");
+
+    out->Resize(framework::make_ddim({num}));
+
+    T* out_data = out->mutable_data<T>(context.GetPlace());
+
+    if (num > 1) {
+      T step = (stop - start) / (num - 1);
+      T value = start;
+      for (int i = 0; i < num; ++i) {
+        out_data[i] = value;
+        value += step;
+      }
+    } else {
+      out_data[0] = start;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/fluid/operators/lod_reset_op.cc b/paddle/fluid/operators/lod_reset_op.cc
index e0ab02cd90cdee848250a6aba882b0cb0c17abd7..458037c5aca6af4c8c97b2da630c35929770c156 100644
--- a/paddle/fluid/operators/lod_reset_op.cc
+++ b/paddle/fluid/operators/lod_reset_op.cc
@@ -30,10 +30,10 @@ class LoDResetOp : public framework::OperatorWithKernel {
 
     if (!ctx->HasInput("Y")) {
       auto level0 = ctx->Attrs().Get<std::vector<int>>("target_lod");
-      PADDLE_ENFORCE_GT(level0.size(), 1,
+      PADDLE_ENFORCE_GT(level0.size(), 0,
                         "If Input(Y) not provided, the target lod should be "
                         "specified by attribute `target_lod`.");
-    } else {
+    } else if (ctx->IsRuntime()) {
       ctx->ShareLoD("Y", "Out");
     }
 
@@ -48,6 +48,23 @@ class LoDResetOp : public framework::OperatorWithKernel {
   }
 };
 
+class LoDResetOpVarTypeInference : public framework::VarTypeInference {
+ public:
+  void operator()(framework::InferVarTypeContext *ctx) const override {
+    auto x_var_name = ctx->Input("X").front();
+    auto out_var_name = ctx->Output("Out").front();
+    if (ctx->HasInput("Y")) {
+      auto y_var_name = ctx->Input("Y").front();
+      auto y_lod_level = std::max(ctx->GetLoDLevel(y_var_name), 1);
+      ctx->SetLoDLevel(out_var_name, y_lod_level);
+    } else {
+      ctx->SetLoDLevel(out_var_name, 1);
+    }
+    ctx->SetDataType(out_var_name, ctx->GetDataType(x_var_name));
+    ctx->SetType(out_var_name, paddle::framework::proto::VarType::LOD_TENSOR);
+  }
+};
+
 class LoDResetOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
@@ -177,9 +194,10 @@ DECLARE_NO_NEED_BUFFER_VARS_INFERENCE(LoDResetGradNoNeedBufferVarInference,
 
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(lod_reset, ops::LoDResetOp, ops::LoDResetOpMaker,
-                  ops::LoDResetGradDescMaker);
+                  ops::LoDResetGradDescMaker, ops::LoDResetOpVarTypeInference);
 REGISTER_OPERATOR(lod_reset_grad, ops::LoDResetGradOp,
                   ops::LoDResetGradNoNeedBufferVarInference);
+
 REGISTER_OP_CPU_KERNEL(
     lod_reset, ops::LoDResetKernel<paddle::platform::CPUPlace, float>,
     ops::LoDResetKernel<paddle::platform::CPUPlace, double>,
diff --git a/paddle/fluid/operators/lod_reset_op.h b/paddle/fluid/operators/lod_reset_op.h
index d36aa0ce025a1c0f717913131fcc75040d16afac..1c2f0b0ac8ab4be35e4716acc7be3f05b9d63805 100644
--- a/paddle/fluid/operators/lod_reset_op.h
+++ b/paddle/fluid/operators/lod_reset_op.h
@@ -63,7 +63,7 @@ class LoDResetKernel : public framework::OpKernel<T> {
                       "Target LoD should be a vector end with the "
                       "first dimension of Input(X).");
     for (size_t i = 0; i < level0.size() - 1; ++i) {
-      PADDLE_ENFORCE(level0[i + 1] > level0[i],
+      PADDLE_ENFORCE(level0[i + 1] >= level0[i],
                      "Target LoD should be an ascending vector.");
     }
 
diff --git a/paddle/fluid/operators/recurrent_op.cc b/paddle/fluid/operators/recurrent_op.cc
index 2898a62ddbac524ceb212cac5f34aeda3b1e01cb..1a2feee11c951cd4a55958df58f3756472f64769 100644
--- a/paddle/fluid/operators/recurrent_op.cc
+++ b/paddle/fluid/operators/recurrent_op.cc
@@ -23,6 +23,7 @@ constexpr char kInitialStates[] = "initial_states";
 constexpr char kParameters[] = "parameters";
 constexpr char kOutputs[] = "outputs";
 constexpr char kStepScopes[] = "step_scopes";
+constexpr char kHasStates[] = "has_states";
 constexpr char kExStates[] = "ex_states";
 constexpr char kStates[] = "states";
 constexpr char kStepBlock[] = "sub_block";
@@ -241,11 +242,16 @@ class RecurrentOp : public RecurrentBase {
  private:
   void RunImpl(const framework::Scope &scope,
                const platform::Place &place) const override {
+    bool has_state = Attr<bool>(kHasStates);
     auto seq_len = static_cast<size_t>(this->GetSequenceLength(scope));
     VLOG(3) << "Static RNN input sequence length = " << seq_len;
     StepScopes scopes = CreateStepScopes(scope, seq_len);
     auto reverse = Attr<bool>(kReverse);
 
+    // get device context from pool
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
+    auto &dev_ctx = *pool.Get(place);
+
     framework::Executor executor(place);
     auto *block = Attr<framework::BlockDesc *>(kStepBlock);
 
@@ -269,15 +275,17 @@ class RecurrentOp : public RecurrentBase {
             inside->Resize(framework::make_ddim(dims));
           });
 
-      if (i == 0) {
-        // Link initial states  --> ex_states
-        LinkTensor(scope, Inputs(kInitialStates), &cur_scope,
-                   Attr<std::vector<std::string>>(kExStates));
-      } else {
-        auto &ex_scope = scopes.ExScope();
-        // Link ex_scope::state --> cur_scope::ex_state
-        LinkTensor(ex_scope, Attr<std::vector<std::string>>(kStates),
-                   &cur_scope, Attr<std::vector<std::string>>(kExStates));
+      if (has_state) {
+        if (i == 0) {
+          // Link initial states  --> ex_states
+          LinkTensor(scope, Inputs(kInitialStates), &cur_scope,
+                     Attr<std::vector<std::string>>(kExStates));
+        } else {
+          auto &ex_scope = scopes.ExScope();
+          // Link ex_scope::state --> cur_scope::ex_state
+          LinkTensor(ex_scope, Attr<std::vector<std::string>>(kStates),
+                     &cur_scope, Attr<std::vector<std::string>>(kExStates));
+        }
       }
 
       // Every inputs are linked now, execute!
@@ -286,11 +294,6 @@ class RecurrentOp : public RecurrentBase {
                    std::vector<std::string>() /*skip_ref_cnt_vars*/,
                    true /*force_disable_gc*/);
 
-      // get device context from pool
-      platform::DeviceContextPool &pool =
-          platform::DeviceContextPool::Instance();
-      auto &dev_ctx = *pool.Get(place);
-
       // Copy inside::output -> outside::output
       //    outside::output[seq_offset: seq_offset + 1] = inside::output
       this->LinkTensorWithCallback(
@@ -333,13 +336,13 @@ class RecurrentGradOp : public RecurrentBase {
  private:
   void RunImpl(const framework::Scope &scope,
                const platform::Place &place) const override {
-    auto seq_len = static_cast<size_t>(GetSequenceLength(scope));
+    bool has_state = Attr<bool>(kHasStates);
+    const size_t seq_len = static_cast<size_t>(GetSequenceLength(scope));
     StepScopes scopes = CreateStepScopes(scope, seq_len);
     auto reverse = Attr<bool>(kReverse);
 
     framework::Executor executor(place);
     auto *block = Attr<framework::BlockDesc *>(kStepBlock);
-
     auto *program = block->Program();
 
     // get device context from pool
@@ -350,6 +353,7 @@ class RecurrentGradOp : public RecurrentBase {
       size_t seq_offset = reverse ? step_id : seq_len - step_id - 1;
       VLOG(3) << "Recurrent backward operate at the time step " << seq_offset;
       auto &cur_scope = scopes.CurScope();
+
       // Link outside::output_grads --> inside::output_grads
       //   inside::output_grad = outside::output_grad[seq_offset:seq_offset+1]
       LinkTensorWithCallback(
@@ -370,30 +374,32 @@ class RecurrentGradOp : public RecurrentBase {
         VLOG(10) << " RNN output gradients = [" << sout.str() << "]";
       }
 
-      // Link states
-      //   if cur_scope::cur_state_grad in out_grads:
-      //     cur_scope::cur_state_grad += ex_scope::ex_state_grad
-      //   else:
-      //     ex_scope::ex_state_grad --> cur_scope::cur_state_grad
-      if (step_id != 0) {  // not at beginning
-        auto &ex_scope = scopes.ExScope();
-        auto ex_state_grads =
-            GradVarLists(Attr<std::vector<std::string>>(kExStates));
-        auto cur_state_grads =
-            GradVarLists(Attr<std::vector<std::string>>(kStates));
-
-        PADDLE_ENFORCE_EQ(ex_state_grads.size(), cur_state_grads.size());
-        for (size_t i = 0; i < ex_state_grads.size(); ++i) {
-          auto &cur_grad = cur_state_grads[i];
-          auto &ex_grad = ex_state_grads[i];
-          auto &ex_tensor =
-              ex_scope.FindVar(ex_grad)->Get<framework::LoDTensor>();
-
-          VLOG(10) << " RNN link " << cur_grad << " from " << ex_grad;
-          auto *cur_grad_var = cur_scope.Var(cur_grad);
-          auto cur_grad_tensor =
-              cur_grad_var->GetMutable<framework::LoDTensor>();
-          framework::TensorCopy(ex_tensor, place, dev_ctx, cur_grad_tensor);
+      if (has_state) {
+        // Link states
+        //   if cur_scope::cur_state_grad in out_grads:
+        //     cur_scope::cur_state_grad += ex_scope::ex_state_grad
+        //   else:
+        //     ex_scope::ex_state_grad --> cur_scope::cur_state_grad
+        if (step_id != 0) {  // not at beginning
+          auto &ex_scope = scopes.ExScope();
+          auto ex_state_grads =
+              GradVarLists(Attr<std::vector<std::string>>(kExStates));
+          auto cur_state_grads =
+              GradVarLists(Attr<std::vector<std::string>>(kStates));
+
+          PADDLE_ENFORCE_EQ(ex_state_grads.size(), cur_state_grads.size());
+          for (size_t i = 0; i < ex_state_grads.size(); ++i) {
+            auto &cur_grad = cur_state_grads[i];
+            auto &ex_grad = ex_state_grads[i];
+            auto &ex_tensor =
+                ex_scope.FindVar(ex_grad)->Get<framework::LoDTensor>();
+
+            VLOG(10) << " RNN link " << cur_grad << " from " << ex_grad;
+            auto *cur_grad_var = cur_scope.Var(cur_grad);
+            auto cur_grad_tensor =
+                cur_grad_var->GetMutable<framework::LoDTensor>();
+            framework::TensorCopy(ex_tensor, place, dev_ctx, cur_grad_tensor);
+          }
         }
       }
 
@@ -442,8 +448,8 @@ class RecurrentGradOp : public RecurrentBase {
           }
 
           auto new_inside_name = cur_scope.Rename(inside_grad_name);
-          // sum gradient
 
+          // sum gradient
           auto sum_op = framework::OpRegistry::CreateOp(
               "sum", {{"X", {pg_names[param_id], new_inside_name}}},
               {{"Out", {pg_names[param_id]}}},
@@ -475,22 +481,33 @@ class RecurrentGradOp : public RecurrentBase {
           true /*is_backward*/);
       VLOG(5) << "Link outside gradient finished ";
 
-      if (step_id + 1 == seq_len) {  // at_end
-        // copy initialize states gradient from inside to outside
-        LinkTensorWithCallback(
-            cur_scope, GradVarLists(Attr<std::vector<std::string>>(kExStates)),
-            scope, Outputs(kInitStateGrads),
-            [&](const framework::LoDTensor &inside,
-                framework::LoDTensor *outside) {
-              outside->Resize(inside.dims());
-              outside->mutable_data(place, inside.type());
-              framework::TensorCopy(inside, place, dev_ctx, outside);
-            },
-            true /*is_backward*/);
-        VLOG(5) << "Link initialize state gradient finished ";
+      if (has_state) {
+        if (step_id + 1 == seq_len) {  // at_end
+          // copy initialize states gradient from inside to outside
+          LinkTensorWithCallback(
+              cur_scope,
+              GradVarLists(Attr<std::vector<std::string>>(kExStates)), scope,
+              Outputs(kInitStateGrads),
+              [&](const framework::LoDTensor &inside,
+                  framework::LoDTensor *outside) {
+                outside->Resize(inside.dims());
+                outside->mutable_data(place, inside.type());
+                framework::TensorCopy(inside, place, dev_ctx, outside);
+              },
+              true /*is_backward*/);
+          VLOG(5) << "Link initialize state gradient finished ";
+        }
       }
       scopes.Next();
     }
+    // Delete the scope of StepScopes
+    dev_ctx.Wait();
+    auto *var = scope.FindVar(Input(kStepScopes));
+    PADDLE_ENFORCE(var != nullptr);
+    auto step_scopes = var->GetMutable<StepScopeVar>();
+    for (auto *sub_scope : *step_scopes) {
+      const_cast<framework::Scope &>(scope).DeleteScope(sub_scope);
+    }
   }
 
  private:
@@ -541,6 +558,7 @@ class RecurrentOpProtoMaker : public framework::OpProtoAndCheckerMaker {
         .AsDuplicable();
     AddOutput(kStepScopes,
               "StepScopes contain all local variables in each time step.");
+    AddAttr<bool>(kHasStates, "Whether has states.").SetDefault(false);
     AddAttr<std::vector<std::string>>(kExStates,
                                       string::Sprintf(
                                           R"DOC(The ex-state variable names.
@@ -624,20 +642,44 @@ class RecurrentGradOpDescMaker : public framework::SingleGradOpDescMaker {
 class RecurrentGradOpShapeInference : public framework::InferShapeBase {
  public:
   void operator()(framework::InferShapeContext *ctx) const override {
-    std::vector<std::string> input{kInputs, kInitialStates};
     std::vector<std::string> output{kOutputs};
-    for (auto &s : input) {
-      // NOTE(zcd): In some case, some of kInputs doesn't have gradient.
-      PADDLE_ENFORCE(ctx->HasInputs(s));
-    }
-    for (auto &s : output) {
-      PADDLE_ENFORCE(ctx->HasInputs(s));
+
+    // In some case the kInitialStates is empty.
+    // If the kInitialStates is empty, all the states should be empty.
+    if (!ctx->HasInputs(kInitialStates)) {
+      PADDLE_ENFORCE_EQ(
+          ctx->Attrs().Get<std::vector<std::string>>(kExStates).size(), 0,
+          "The Attr(%s) should be empty.", kExStates);
+      PADDLE_ENFORCE_EQ(
+          ctx->Attrs().Get<std::vector<std::string>>(kStates).size(), 0,
+          "The Attr(%s) should be empty.", kStates);
     }
-    for (auto &s : input) {
-      ctx->SetOutputsDim(framework::GradVarName(s), ctx->GetInputsDim(s));
+
+    PADDLE_ENFORCE(ctx->HasInputs(kInputs),
+                   "The input(%s) should not be empty.", kInputs);
+    PADDLE_ENFORCE(ctx->HasInputs(kOutputs),
+                   "The input(%s) should not be empty.", kOutputs);
+
+    // In some case the kInitialStates is empty.
+    if (ctx->HasInputs(kInitialStates)) {
+      PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName(kInitialStates)),
+                     "The output of(%s) should not be empty.",
+                     framework::GradVarName(kInitialStates));
+      ctx->SetOutputsDim(framework::GradVarName(kInitialStates),
+                         ctx->GetInputsDim(kInitialStates));
     }
+
+    PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName(kInputs)),
+                   "The output of(%s) should not be empty.",
+                   framework::GradVarName(kInputs));
+    ctx->SetOutputsDim(framework::GradVarName(kInputs),
+                       ctx->GetInputsDim(kInputs));
+
+    // In some case the kParameters is empty.
     if (ctx->HasInputs(kParameters)) {
-      PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName(kParameters)));
+      PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName(kParameters)),
+                     "The output of(%s) should not be empty.",
+                     framework::GradVarName(kParameters));
       ctx->SetOutputsDim(framework::GradVarName(kParameters),
                          ctx->GetInputsDim(kParameters));
     }
diff --git a/paddle/fluid/operators/reduce_ops/reduce_all_op.cc b/paddle/fluid/operators/reduce_ops/reduce_all_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..b087fbbb94c7ba2f7449f6bda56010dee1c38ea6
--- /dev/null
+++ b/paddle/fluid/operators/reduce_ops/reduce_all_op.cc
@@ -0,0 +1,20 @@
+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/operators/reduce_ops/reduce_all_op.h"
+
+REGISTER_REDUCE_OP(reduce_all);
+REGISTER_OP_CPU_KERNEL(reduce_all,
+                       ops::ReduceKernel<paddle::platform::CPUDeviceContext,
+                                         bool, ops::AllFunctor>);
diff --git a/paddle/fluid/operators/reduce_ops/reduce_all_op.cu b/paddle/fluid/operators/reduce_ops/reduce_all_op.cu
new file mode 100644
index 0000000000000000000000000000000000000000..bd94ba263d957d0d65506ecd802bf43add6e2fb4
--- /dev/null
+++ b/paddle/fluid/operators/reduce_ops/reduce_all_op.cu
@@ -0,0 +1,19 @@
+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/operators/reduce_ops/reduce_all_op.h"
+
+REGISTER_OP_CUDA_KERNEL(reduce_all,
+                        ops::ReduceKernel<paddle::platform::CUDADeviceContext,
+                                          bool, ops::AllFunctor>);
diff --git a/paddle/fluid/operators/reduce_ops/reduce_all_op.h b/paddle/fluid/operators/reduce_ops/reduce_all_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..ba159dd703c8904784546eda262bf7be77967d48
--- /dev/null
+++ b/paddle/fluid/operators/reduce_ops/reduce_all_op.h
@@ -0,0 +1,29 @@
+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+#pragma once
+
+#include "paddle/fluid/operators/reduce_ops/reduce_op.h"
+
+namespace paddle {
+namespace operators {
+
+struct AllFunctor {
+  template <typename DeviceContext, typename X, typename Y, typename Dim>
+  void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) {
+    y->device(place) = x->all(dim);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/fluid/operators/reduce_ops/reduce_any_op.cc b/paddle/fluid/operators/reduce_ops/reduce_any_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..d865dcb3c935b76b8da25d723a5f780fb4de255b
--- /dev/null
+++ b/paddle/fluid/operators/reduce_ops/reduce_any_op.cc
@@ -0,0 +1,20 @@
+// Copyright (c) 2018 PaddlePaddle Authors. Any Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/operators/reduce_ops/reduce_any_op.h"
+
+REGISTER_REDUCE_OP(reduce_any);
+REGISTER_OP_CPU_KERNEL(reduce_any,
+                       ops::ReduceKernel<paddle::platform::CPUDeviceContext,
+                                         bool, ops::AnyFunctor>);
diff --git a/paddle/fluid/operators/reduce_ops/reduce_any_op.cu b/paddle/fluid/operators/reduce_ops/reduce_any_op.cu
new file mode 100644
index 0000000000000000000000000000000000000000..66f0c9997ea1e27cf172a6839a68d2eb23395c4d
--- /dev/null
+++ b/paddle/fluid/operators/reduce_ops/reduce_any_op.cu
@@ -0,0 +1,19 @@
+// Copyright (c) 2018 PaddlePaddle Authors. Any Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/operators/reduce_ops/reduce_any_op.h"
+
+REGISTER_OP_CUDA_KERNEL(reduce_any,
+                        ops::ReduceKernel<paddle::platform::CUDADeviceContext,
+                                          bool, ops::AnyFunctor>);
diff --git a/paddle/fluid/operators/reduce_ops/reduce_any_op.h b/paddle/fluid/operators/reduce_ops/reduce_any_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..b36bad9cada259932d2bd77c2426fbb46790de76
--- /dev/null
+++ b/paddle/fluid/operators/reduce_ops/reduce_any_op.h
@@ -0,0 +1,29 @@
+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+#pragma once
+
+#include "paddle/fluid/operators/reduce_ops/reduce_op.h"
+
+namespace paddle {
+namespace operators {
+
+struct AnyFunctor {
+  template <typename DeviceContext, typename X, typename Y, typename Dim>
+  void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) {
+    y->device(place) = x->any(dim);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/fluid/operators/rnn_memory_helper_op.cc b/paddle/fluid/operators/rnn_memory_helper_op.cc
index 834dd1eabd68db6c8b571071f8043589c66f8671..b00cc07dea920a6d7caa8b70c99d84b72a785a99 100644
--- a/paddle/fluid/operators/rnn_memory_helper_op.cc
+++ b/paddle/fluid/operators/rnn_memory_helper_op.cc
@@ -40,9 +40,12 @@ class RNNMemoryHelperOp : public framework::OperatorBase {
                    "Cannot find out_var in scope, out_var_name is %s",
                    out_name);
 
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
+    auto &dev_ctx = *pool.Get(dev_place);
+
     auto *out_tensor = out_var->GetMutable<framework::LoDTensor>();
     auto &mem_tensor = mem_var->Get<framework::LoDTensor>();
-    framework::TensorCopySync(mem_tensor, dev_place, out_tensor);
+    framework::TensorCopy(mem_tensor, dev_place, dev_ctx, out_tensor);
     out_tensor->set_lod(mem_tensor.lod());
   }
 };
@@ -92,6 +95,9 @@ class RNNMemoryHelperGradOp : public framework::OperatorBase {
                    "Cannot find in_grad_var in scope, name is %s",
                    in_grad_var_name);
 
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
+    auto &dev_ctx = *pool.Get(dev_place);
+
     if (out_grad_var == nullptr) {
       VLOG(5) << "Using fill constant 0 as starting gradient";
       auto in_var_name = Input("X");
@@ -109,7 +115,8 @@ class RNNMemoryHelperGradOp : public framework::OperatorBase {
     } else {
       auto &out_grad_tensor = out_grad_var->Get<framework::LoDTensor>();
       auto *in_grad_tensor = in_grad_var->GetMutable<framework::LoDTensor>();
-      framework::TensorCopySync(out_grad_tensor, dev_place, in_grad_tensor);
+      framework::TensorCopy(out_grad_tensor, dev_place, dev_ctx,
+                            in_grad_tensor);
       in_grad_tensor->set_lod(out_grad_tensor.lod());
     }
   }
diff --git a/paddle/fluid/operators/squared_l2_distance_op.h b/paddle/fluid/operators/squared_l2_distance_op.h
index e0133d33e6a840d2d06832393a064df978cb9cbc..12a8f05b5a603417ead8ebd250ff7951f928f4a1 100644
--- a/paddle/fluid/operators/squared_l2_distance_op.h
+++ b/paddle/fluid/operators/squared_l2_distance_op.h
@@ -77,6 +77,9 @@ class SquaredL2DistanceGradKernel : public framework::OpKernel<T> {
     auto* x_g = context.Output<Tensor>(framework::GradVarName("X"));
     auto* y_g = context.Output<Tensor>(framework::GradVarName("Y"));
 
+    PADDLE_ENFORCE_NOT_NULL(x_g);
+    PADDLE_ENFORCE_NOT_NULL(y_g);
+
     auto sub_result = EigenMatrix<T>::From(*in0);
     auto out_grad = EigenMatrix<T>::From(*in1);
 
@@ -92,31 +95,28 @@ class SquaredL2DistanceGradKernel : public framework::OpKernel<T> {
     // propagate back to input
     auto& eigen_place =
         *context.template device_context<DeviceContext>().eigen_device();
-    if (x_g) {
-      x_g->mutable_data<T>(context.GetPlace());
-      // eigen matrix
-      auto x_grad =
-          EigenMatrix<T>::From(*x_g, framework::make_ddim({x_dims[0], cols}));
-      // dimensions are same with subResult
-      x_grad.device(eigen_place) = grad_mat;
-    }
 
-    if (y_g) {
-      y_g->mutable_data<T>(context.GetPlace());
-
-      PADDLE_ENFORCE_GE(sub_result.dimensions()[0], y_dims[0],
-                        "First dimension of gradient must be greater or "
-                        "equal than first dimension of target.");
-
-      if (sub_result.dimensions()[0] == y_dims[0]) {
-        auto y_grad =
-            EigenMatrix<T>::From(*y_g, framework::make_ddim({y_dims[0], cols}));
-        y_grad.device(eigen_place) = -1 * grad_mat;
-      } else {
-        auto col_sum_res = -1 * (grad_mat.sum(Eigen::array<int, 1>({{0}})));
-        auto y_grad = EigenVector<T>::Flatten(*y_g);
-        y_grad.device(eigen_place) = col_sum_res;
-      }
+    x_g->mutable_data<T>(context.GetPlace());
+    // eigen matrix
+    auto x_grad =
+        EigenMatrix<T>::From(*x_g, framework::make_ddim({x_dims[0], cols}));
+    // dimensions are same with subResult
+    x_grad.device(eigen_place) = grad_mat;
+
+    y_g->mutable_data<T>(context.GetPlace());
+
+    PADDLE_ENFORCE_GE(sub_result.dimensions()[0], y_dims[0],
+                      "First dimension of gradient must be greater or "
+                      "equal than first dimension of target.");
+
+    if (sub_result.dimensions()[0] == y_dims[0]) {
+      auto y_grad =
+          EigenMatrix<T>::From(*y_g, framework::make_ddim({y_dims[0], cols}));
+      y_grad.device(eigen_place) = -1 * grad_mat;
+    } else {
+      auto col_sum_res = -1 * (grad_mat.sum(Eigen::array<int, 1>({{0}})));
+      auto y_grad = EigenVector<T>::Flatten(*y_g);
+      y_grad.device(eigen_place) = col_sum_res;
     }
   }
 };
diff --git a/paddle/fluid/platform/lodtensor_printer.cc b/paddle/fluid/platform/lodtensor_printer.cc
index a5aa1a4148686b032c52f99497252fde4867438f..07eaf42d2d3bc20e7f7dc56bb0f4e0cc2fbac5e3 100644
--- a/paddle/fluid/platform/lodtensor_printer.cc
+++ b/paddle/fluid/platform/lodtensor_printer.cc
@@ -52,16 +52,26 @@ void PrintVar(framework::Scope* scope, const std::string& var_name,
     return;
   }
 
-#define PrintLoDTensorCallback(cpp_type, proto_type)             \
-  do {                                                           \
-    if (tensor->type() == proto_type) {                          \
-      print_lod_tensor<cpp_type>(var_name, *tensor, print_info); \
-      return;                                                    \
-    }                                                            \
+  framework::LoDTensor printed_tensor;
+  printed_tensor.set_lod(tensor->lod());
+  printed_tensor.Resize(tensor->dims());
+  if (platform::is_cpu_place(tensor->place())) {
+    printed_tensor.ShareDataWith(*tensor);
+  } else {
+    platform::CPUPlace place;
+    framework::TensorCopy(*tensor, place, &printed_tensor);
+  }
+
+#define PrintLoDTensorCallback(cpp_type, proto_type)                    \
+  do {                                                                  \
+    if (tensor->type() == proto_type) {                                 \
+      print_lod_tensor<cpp_type>(var_name, printed_tensor, print_info); \
+      return;                                                           \
+    }                                                                   \
   } while (0)
 
   _ForEachDataType_(PrintLoDTensorCallback);
-  VLOG(1) << "PrintVar: unrecognized data type:" << tensor->type();
+  VLOG(1) << "PrintVar: unrecognized data type:" << printed_tensor.type();
 }
 
 }  // end namespace platform
diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc
index f0ea6d9b0a751c86e3911c35d9403a32604056d7..a8a2a94d473b18fdcd78771063ef4565c7fe0e42 100644
--- a/paddle/fluid/pybind/pybind.cc
+++ b/paddle/fluid/pybind/pybind.cc
@@ -1366,6 +1366,10 @@ All parameter, weight, gradient are variables in Paddle.
           "cache_runtime_context",
           [](const BuildStrategy &self) { return self.cache_runtime_context_; },
           [](BuildStrategy &self, bool b) { self.cache_runtime_context_ = b; })
+      .def_property(
+          "cache_expected_kernel",
+          [](const BuildStrategy &self) { return self.cache_expected_kernel_; },
+          [](BuildStrategy &self, bool b) { self.cache_expected_kernel_ = b; })
       .def("_finalize_strategy_and_create_passes",
            [](BuildStrategy &self) -> std::shared_ptr<ir::PassBuilder> {
              return self.CreatePassesFromStrategy(true);
diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh
index 43db744b5edeb48f87c4413961e6ded3bd45f86d..7bb713493182239b2fd17f7b7fb496afdc9b8e6c 100755
--- a/paddle/scripts/paddle_build.sh
+++ b/paddle/scripts/paddle_build.sh
@@ -202,6 +202,7 @@ function cmake_gen() {
         -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
         -DWITH_CONTRIB=${WITH_CONTRIB:-ON}
         -DWITH_INFERENCE_API_TEST=${WITH_INFERENCE_API_TEST:-ON}
+        -DWITH_HIGH_LEVEL_API_TEST=${WITH_HIGH_LEVEL_API_TEST:-OFF}
         -DINFERENCE_DEMO_INSTALL_DIR=${INFERENCE_DEMO_INSTALL_DIR}
         -DWITH_ANAKIN=${WITH_ANAKIN:-OFF}
         -DANAKIN_BUILD_FAT_BIN=${ANAKIN_BUILD_FAT_BIN:OFF}
@@ -234,6 +235,7 @@ EOF
         -DCMAKE_EXPORT_COMPILE_COMMANDS=ON \
         -DWITH_CONTRIB=${WITH_CONTRIB:-ON} \
         -DWITH_INFERENCE_API_TEST=${WITH_INFERENCE_API_TEST:-ON} \
+        -DWITH_HIGH_LEVEL_API_TEST=${WITH_HIGH_LEVEL_API_TEST:-OFF} \
         -DINFERENCE_DEMO_INSTALL_DIR=${INFERENCE_DEMO_INSTALL_DIR} \
         -DWITH_ANAKIN=${WITH_ANAKIN:-OFF} \
         -DANAKIN_BUILD_FAT_BIN=${ANAKIN_BUILD_FAT_BIN:OFF}\
diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py
index 983d8243b1d8aa6c8d01855d6dbeab76c335f70c..3dc2b0c895116155f41df3ca66125fff3ede5ead 100644
--- a/python/paddle/fluid/__init__.py
+++ b/python/paddle/fluid/__init__.py
@@ -175,6 +175,7 @@ def __bootstrap__():
         read_env_flags.append('communicator_thread_pool_size')
         read_env_flags.append('communicator_max_merge_var_num')
         read_env_flags.append('communicator_fake_rpc')
+        read_env_flags.append('communicator_send_wait_times')
         if core.is_compiled_with_brpc():
             read_env_flags.append('max_body_size')
             #set brpc max body size
diff --git a/python/paddle/fluid/dataset.py b/python/paddle/fluid/dataset.py
index e655fd4a976a8a6fa2811ddc43de3d1f231229d5..1a023f61675ed62c141bb6e71fabbdf0086b0c64 100644
--- a/python/paddle/fluid/dataset.py
+++ b/python/paddle/fluid/dataset.py
@@ -136,6 +136,7 @@ class DatasetBase(object):
             slot_var.name = var.name
             if var.lod_level == 0:
                 slot_var.is_dense = True
+                slot_var.shape.extend(var.shape)
             if var.dtype == core.VarDesc.VarType.FP32:
                 slot_var.type = "float"
             elif var.dtype == core.VarDesc.VarType.INT64:
diff --git a/python/paddle/fluid/device_worker.py b/python/paddle/fluid/device_worker.py
index 7fc72191884020f4cc57c9269b636161635f06d0..0998f779acfea23f3a494a25b43a6fa824b985f1 100644
--- a/python/paddle/fluid/device_worker.py
+++ b/python/paddle/fluid/device_worker.py
@@ -26,8 +26,8 @@ class DeviceWorker(object):
         """
         Init.
         """
-        self.program_ = None
-        self.infer_ = None
+        self._program = None
+        self._infer = None
 
     def _set_infer(self, infer=False):
         """
@@ -36,7 +36,7 @@ class DeviceWorker(object):
         Args:
             infer(bool): whether to do inference
         """
-        self.infer_ = infer
+        self._infer = infer
 
     def _set_fleet_desc(self, fleet_desc):
         """
@@ -45,7 +45,7 @@ class DeviceWorker(object):
         Args:
             fleet_desc(PSParameter): pslib.PSParameter object
         """
-        self.fleet_desc_ = fleet_desc
+        self._fleet_desc = fleet_desc
 
     def _set_program(self, program):
         """
@@ -54,7 +54,7 @@ class DeviceWorker(object):
         Args:
             program(Program): a Program object
         """
-        self.program_ = program
+        self._program = program
 
     def _gen_worker_desc(self, trainer_desc):
         """
@@ -88,7 +88,7 @@ class Hogwild(DeviceWorker):
             trainer_desc(TrainerDesc): a TrainerDesc object
         """
         trainer_desc.device_worker_name = "HogwildWorker"
-        if self.infer_:
+        if self._infer:
             # just ignore feed op for inference model
             trainer_desc.hogwild_param.skip_ops.extend(["feed"])
 
@@ -113,11 +113,11 @@ class DownpourSGD(DeviceWorker):
             trainer_desc(TrainerDesc): a TrainerDesc object
         """
         dense_table_set = set()
-        program_id = str(id(self.program_))
-        if self.program_ == None:
+        program_id = str(id(self._program))
+        if self._program == None:
             print("program of current device worker is not configured")
             exit(-1)
-        opt_info = self.program_._fleet_opt
+        opt_info = self._program._fleet_opt
         program_configs = opt_info["program_configs"]
         downpour = trainer_desc.downpour_param
 
@@ -140,7 +140,7 @@ class DownpourSGD(DeviceWorker):
         trainer_desc.device_worker_name = "DownpourWorker"
         pull_thread = trainer_desc.pull_dense_param
         pull_thread.device_num = trainer_desc.thread_num
-        for i in self.fleet_desc_.trainer_param.dense_table:
+        for i in self._fleet_desc.trainer_param.dense_table:
             if i.table_id in dense_table_set:
                 dense_table = pull_thread.dense_table.add()
                 dense_table.dense_value_name.extend(i.dense_variable_name)
@@ -148,29 +148,29 @@ class DownpourSGD(DeviceWorker):
                     i.table_id
         sparse_table = downpour.sparse_table.add()
         sparse_table.table_id = \
-                    self.fleet_desc_.trainer_param.sparse_table[0].table_id
+                    self._fleet_desc.trainer_param.sparse_table[0].table_id
         sparse_table.sparse_key_name.extend(
-            self.fleet_desc_.trainer_param.sparse_table[0].slot_key)
+            self._fleet_desc.trainer_param.sparse_table[0].slot_key)
         sparse_table.sparse_value_name.extend(
-            self.fleet_desc_.trainer_param.sparse_table[0].slot_value)
+            self._fleet_desc.trainer_param.sparse_table[0].slot_value)
         sparse_table.sparse_grad_name.extend(
-            self.fleet_desc_.trainer_param.sparse_table[0].slot_gradient)
+            self._fleet_desc.trainer_param.sparse_table[0].slot_gradient)
         sparse_table.emb_dim = \
-                    self.fleet_desc_.server_param.downpour_server_param.downpour_table_param[
+                    self._fleet_desc.server_param.downpour_server_param.downpour_table_param[
                         0].accessor.fea_dim - 2
         sparse_table.fea_dim = sparse_table.emb_dim + 2
         # TODO(guru4elephant): hard code here, need to improve
         sparse_table.label_var_name = "click"
 
-        for i in self.fleet_desc_.trainer_param.dense_table:
+        for i in self._fleet_desc.trainer_param.dense_table:
             if i.table_id in dense_table_set:
                 dense_table = downpour.dense_table.add()
                 dense_table.table_id = i.table_id
                 dense_table.dense_value_name.extend(i.dense_variable_name)
                 dense_table.dense_grad_name.extend(
                     i.dense_gradient_variable_name)
-                downpour.skip_ops.extend(self.fleet_desc_.trainer_param.skip_op)
-        if self.infer_:
+                downpour.skip_ops.extend(self._fleet_desc.trainer_param.skip_op)
+        if self._infer:
             downpour.push_dense = False
             downpour.push_sparse = False
 
diff --git a/python/paddle/fluid/executor.py b/python/paddle/fluid/executor.py
index e15197037e1d901855883919b02a1574b7bc9a29..fa8b49a021294e8555e979459615b1956d9b2b55 100644
--- a/python/paddle/fluid/executor.py
+++ b/python/paddle/fluid/executor.py
@@ -712,10 +712,6 @@ class Executor(object):
         if dataset == None:
             raise RuntimeError("dataset is needed and should be initialized")
 
-        if not isinstance(self.place, core.CPUPlace):
-            raise RuntimeError("infer_from_dataset is verified on CPUPlace"
-                               "We will open CUDAPlace in the future")
-
         scope, trainer = self._prepare_trainer(
             program=program,
             dataset=dataset,
@@ -796,10 +792,6 @@ class Executor(object):
         if dataset == None:
             raise RuntimeError("dataset is need and should be initialized")
 
-        if not isinstance(self.place, core.CPUPlace):
-            raise RuntimeError("train_from_dataset is verified on CPUPlace"
-                               "We will open CUDAPlace in the future")
-
         scope, trainer = self._prepare_trainer(
             program=program,
             dataset=dataset,
diff --git a/python/paddle/fluid/incubate/fleet/base/role_maker.py b/python/paddle/fluid/incubate/fleet/base/role_maker.py
index 528f7b3269eb90435d88cffadfa185cc664e430a..bf50a5815dbe445a7a44fd9199ed51f632ff4997 100644
--- a/python/paddle/fluid/incubate/fleet/base/role_maker.py
+++ b/python/paddle/fluid/incubate/fleet/base/role_maker.py
@@ -23,10 +23,10 @@ class RoleMakerBase(object):
     """
 
     def __init__(self):
-        self.role_maker_name_ = ""
-        self.trainer_endpoints_ = []
-        self.pserver_endpoints_ = []
-        self.role_is_generated_ = False
+        self._role_maker_name = ""
+        self._trainer_endpoints = []
+        self._pserver_endpoints = []
+        self._role_is_generated = False
 
     def _is_worker(self):
         """
@@ -45,20 +45,20 @@ class RoleMakerBase(object):
         return get local ip
         """
         import socket
-        self.ip_ = socket.gethostbyname(socket.gethostname())
-        return self.ip_
+        self._ip = socket.gethostbyname(socket.gethostname())
+        return self._ip
 
     def _get_trainer_endpoints(self):
         """
         return trainer endpoints
         """
-        return self.trainer_endpoints_
+        return self._trainer_endpoints
 
     def _get_pserver_endpoints(self):
         """
         return pserver endpoints
         """
-        return self.pserver_endpoints_
+        return self._pserver_endpoints
 
     def _generate_role(self):
         """
@@ -76,59 +76,59 @@ class MPIRoleMaker(RoleMakerBase):
     def __init__(self):
         super(MPIRoleMaker, self).__init__()
         from mpi4py import MPI
-        self.comm_ = MPI.COMM_WORLD
+        self._comm = MPI.COMM_WORLD
         self.MPI = MPI
-        self.ips_ = None
+        self._ips = None
 
     def _get_rank(self):
         """
         return rank
         """
-        self.rank_ = self.comm_.Get_rank()
-        return self.rank_
+        self._rank = self._comm.Get_rank()
+        return self._rank
 
     def _get_size(self):
         """
         return size
         """
-        self.size_ = self.comm_.Get_size()
-        return self.size_
+        self._size = self._comm.Get_size()
+        return self._size
 
     def _all_gather(self, obj):
         """
         all_gather(obj) will call MPI's allgather function
         """
         self._barrier_all()
-        return self.comm_.allgather(obj)
+        return self._comm.allgather(obj)
 
     def _worker_gather(self, obj):
         """
         worker_gather(obj) will call MPI's allgather function
         """
         if self._is_worker():
-            self.node_type_comm_.barrier()
-            return self.node_type_comm_.allgather(obj)
+            self._node_type_comm.barrier()
+            return self._node_type_comm.allgather(obj)
         return None
 
     def _barrier_all(self):
         """
         barrier_all() will call MPI's barrier_all function
         """
-        self.comm_.barrier()
+        self._comm.barrier()
 
     def _get_ips(self):
         """
         collect current distributed job's ip list
         """
-        if self.ips_ == None:
-            self.ips_ = self.comm_.allgather(self._get_local_ip())
-        return self.ips_
+        if self._ips == None:
+            self._ips = self._comm.allgather(self._get_local_ip())
+        return self._ips
 
     def _finalize(self):
         """
         finalize the current MPI instance.
         """
-        self.comm_.finalize()
+        self._comm.finalize()
 
 
 class MPISymetricRoleMaker(MPIRoleMaker):
@@ -140,11 +140,11 @@ class MPISymetricRoleMaker(MPIRoleMaker):
 
     def __init__(self):
         super(MPISymetricRoleMaker, self).__init__()
-        self.node_type_ = None
-        self.proc_per_node_ = 2
+        self._node_type = None
+        self._proc_per_node = 2
 
     def _check_role_generation(self):
-        if not self.role_is_generated_:
+        if not self._role_is_generated:
             sys.stderr.write("generate_role() should be called first")
             sys.exit(-1)
             return False
@@ -163,7 +163,7 @@ class MPISymetricRoleMaker(MPIRoleMaker):
         return whether current process is worker assigned by role maker
         """
         if self._check_role_generation():
-            return self.node_type_ == 1
+            return self._node_type == 1
         return False
 
     def _is_server(self):
@@ -171,7 +171,7 @@ class MPISymetricRoleMaker(MPIRoleMaker):
         return whether current process is server assigned by role maker
         """
         if self._check_role_generation():
-            return self.node_type_ == 0
+            return self._node_type == 0
         return False
 
     def _worker_num(self):
@@ -197,7 +197,7 @@ class MPISymetricRoleMaker(MPIRoleMaker):
         return the index of worker
         """
         if self._check_role_generation():
-            return self.rank_ / self.proc_per_node_
+            return self._rank / self._proc_per_node
         return 0
 
     def _server_index(self):
@@ -205,7 +205,7 @@ class MPISymetricRoleMaker(MPIRoleMaker):
         return the index of server
         """
         if self._check_role_generation():
-            return self.rank_ / self.proc_per_node_
+            return self._rank / self._proc_per_node
         return 0
 
     def _barrier_worker(self):
@@ -214,7 +214,7 @@ class MPISymetricRoleMaker(MPIRoleMaker):
         """
         if self._check_role_generation():
             if self._is_worker():
-                self.node_type_comm_.barrier()
+                self._node_type_comm.barrier()
 
     def _barrier_server(self):
         """
@@ -222,20 +222,20 @@ class MPISymetricRoleMaker(MPIRoleMaker):
         """
         if self._check_role_generation():
             if self._is_server():
-                self.node_type_comm_.barrier()
+                self._node_type_comm.barrier()
 
     def _generate_role(self):
         """
         generate currently process's role
         """
-        if not self.role_is_generated_:
+        if not self._role_is_generated:
             # TODO(guru4elephant): only allow to be called once
-            self.trainer_endpoints_ = self._get_ips()
-            self.pserver_endpoints_ = self._get_ips()
+            self._trainer_endpoints = self._get_ips()
+            self._pserver_endpoints = self._get_ips()
 
-            if 0 == self._get_rank() % self.proc_per_node_ % 2:
-                self.node_type_ = 0
+            if 0 == self._get_rank() % self._proc_per_node % 2:
+                self._node_type = 0
             else:
-                self.node_type_ = 1
-            self.node_type_comm_ = self.comm_.Split(self.node_type_)
-            self.role_is_generated_ = True
+                self._node_type = 1
+            self._node_type_comm = self._comm.Split(self._node_type)
+            self._role_is_generated = True
diff --git a/python/paddle/fluid/incubate/fleet/parameter_server/__init__.py b/python/paddle/fluid/incubate/fleet/parameter_server/__init__.py
index 9b1ec412c731a4b59d0da8847e91e30d8e1d864a..58f17f95323236cfc559a5cb05a3b09bef3b0a5e 100644
--- a/python/paddle/fluid/incubate/fleet/parameter_server/__init__.py
+++ b/python/paddle/fluid/incubate/fleet/parameter_server/__init__.py
@@ -64,9 +64,9 @@ class Fleet(object):
 
     def __init__(self):
         self._opt_info = None  # for fleet only
-        self.role_maker_ = None
-        self.local_ip_ = 0
-        self.is_initialized_ = False
+        self._role_maker = None
+        self._local_ip = 0
+        self._is_initialized = False
 
     def init(self):
         # TODO(guru4elephant)
@@ -78,22 +78,22 @@ class Fleet(object):
             current node's role, e.g. worker, server, etc.
         """
         if not self.is_initialized_:
-            self.role_maker_ = MPISymetricRoleMaker()
-            self.role_maker_._generate_role()
+            self._role_maker = MPISymetricRoleMaker()
+            self._role_maker._generate_role()
             self._fleet_ptr = fluid.core.Fleet()
-            self.is_initialized_ = True
+            self._is_initialized = True
 
     def stop(self):
         """
         stop(): will be called after a user finishes his/her training task. Fleet instance will be
             destroyed when stop() is called.
         """
-        self.role_maker_._barrier_worker()
-        if self.role_maker_._is_first_worker():
+        self._role_maker._barrier_worker()
+        if self._role_maker._is_first_worker():
             self._fleet_ptr.stop_server()
-        self.role_maker_._barrier_worker()
-        self.role_maker_._barrier_all()
-        self.role_maker_._finalize()
+        self._role_maker._barrier_worker()
+        self._role_maker._barrier_all()
+        self._role_maker._finalize()
 
     def init_pserver(self):
         """
@@ -110,15 +110,15 @@ class Fleet(object):
                 sys.exit(-1)
             self._fleet_ptr.init_server(self._dist_desc_str,
                                         self.role_maker_._get_rank())
-            self.local_ip_ = self._fleet_ptr.run_server()
+            self._local_ip = self._fleet_ptr.run_server()
             # barrier_all for init_server
-            self.role_maker_._barrier_all()
-            self.all_ips_ = self.role_maker_._all_gather(self.local_ip_)
+            self._role_maker._barrier_all()
+            self._all_ips = self._role_maker._all_gather(self.local_ip_)
 
-            self._fleet_ptr.gather_servers(self.all_ips_,
-                                           self.role_maker_._get_size())
+            self._fleet_ptr.gather_servers(self._all_ips,
+                                           self._role_maker._get_size())
             # barrier_all for init_worker, wait all workers start
-            self.role_maker_._barrier_all()
+            self._role_maker._barrier_all()
         else:
             print("You should run DistributedOptimizer.minimize() first")
             sys.exit(-1)
@@ -151,21 +151,21 @@ class Fleet(object):
                 print("You should run DistributedOptimizer.minimize() first")
                 sys.exit(-1)
             # barrier_all for init_server, wait for server starts
-            self.role_maker_._barrier_all()
-            self.all_ips_ = self.role_maker_._all_gather(self.local_ip_)
-            self._fleet_ptr.init_worker(self._dist_desc_str, self.all_ips_,
-                                        self.role_maker_._get_size(),
-                                        self.role_maker_._get_rank())
+            self._role_maker._barrier_all()
+            self._all_ips = self._role_maker._all_gather(self.local_ip_)
+            self._fleet_ptr.init_worker(self._dist_desc_str, self._all_ips,
+                                        self._role_maker._get_size(),
+                                        self._role_maker._get_rank())
             # barrier_all for init_worker
-            self.role_maker_._barrier_all()
+            self._role_maker._barrier_all()
             # prepare for client to client communication
             info = self._fleet_ptr.get_clients_info()
-            all_info = self.role_maker_._worker_gather(info[0])
+            all_info = self._role_maker._worker_gather(info[0])
             self._fleet_ptr.gather_clients(all_info)
             self._fleet_ptr.create_client2client_connection()
             # barrier for init model
-            self.role_maker_._barrier_worker()
-            if self.role_maker_._is_first_worker():
+            self._role_maker._barrier_worker()
+            if self._role_maker._is_first_worker():
                 tables = self._dist_desc.trainer_param.dense_table
                 for prog, scope in zip(programs, scopes):
                     prog_id = str(id(prog))
@@ -192,7 +192,7 @@ class Fleet(object):
                                                    int(table.table_id),
                                                    var_name_list)
             # barrier for init model done
-            self.role_maker_._barrier_worker()
+            self._role_maker._barrier_worker()
         else:
             print("You should run DistributedOptimizer.minimize() first")
             sys.exit(-1)
@@ -201,39 +201,39 @@ class Fleet(object):
         """
         return the number of current job's worker num
         """
-        return self.role_maker_._worker_num()
+        return self._role_maker._worker_num()
 
     def get_server_num(self):
         """
         return the number of current job's server num
         """
-        return self.role_maker_._server_num()
+        return self._role_maker._server_num()
 
     def get_worker_index(self):
         """
         return the mpi rank of current worker
         """
-        return self.role_maker_._worker_index()
+        return self._role_maker._worker_index()
 
     def is_worker(self):
         """
         return whether current node is a worker
         """
-        return self.role_maker_._is_worker()
+        return self._role_maker._is_worker()
 
     def is_server(self):
         """
         return whether current node is pserver
         """
-        return self.role_maker_._is_server()
+        return self._role_maker._is_server()
 
     def init_pserver_model(self):
         """
         init pserver model called from pserver
         """
-        if self.role_maker_._is_first_worker():
+        if self._role_maker._is_first_worker():
             self._fleet_ptr.init_model()
-        self.role_maker_._barrier_worker()
+        self._role_maker._barrier_worker()
 
     def save_pserver_model(self, save_path):
         """
diff --git a/python/paddle/fluid/incubate/fleet/parameter_server/node.py b/python/paddle/fluid/incubate/fleet/parameter_server/node.py
index 60035b6e8da3e40158f8be0bafdd911f6bd6f543..641c294c4a6edeb3d9823b4152b0ea158c8faa80 100644
--- a/python/paddle/fluid/incubate/fleet/parameter_server/node.py
+++ b/python/paddle/fluid/incubate/fleet/parameter_server/node.py
@@ -42,13 +42,13 @@ class DownpourServer(Server):
     """
 
     def __init__(self):
-        self.server_ = pslib.ServerParameter()
-        self.server_.downpour_server_param.service_param.start_server_port = 0
-        self.server_.downpour_server_param.service_param.server_class = "DownpourBrpcPsServer"
-        self.server_.downpour_server_param.service_param.client_class = "DownpourBrpcPsClient"
-        self.server_.downpour_server_param.service_param.service_class = "DownpourPsService"
-        self.server_.downpour_server_param.service_param.start_server_port = 0
-        self.server_.downpour_server_param.service_param.server_thread_num = 12
+        self._server = pslib.ServerParameter()
+        self._server.downpour_server_param.service_param.start_server_port = 0
+        self._server.downpour_server_param.service_param.server_class = "DownpourBrpcPsServer"
+        self._server.downpour_server_param.service_param.client_class = "DownpourBrpcPsClient"
+        self._server.downpour_server_param.service_param.service_class = "DownpourPsService"
+        self._server.downpour_server_param.service_param.start_server_port = 0
+        self._server.downpour_server_param.service_param.server_thread_num = 12
 
     def add_sparse_table(self, table_id, learning_rate, slot_key_vars,
                          slot_value_var):
@@ -62,7 +62,7 @@ class DownpourServer(Server):
         Returns:
             return None 
         """
-        table = self.server_.downpour_server_param.downpour_table_param.add()
+        table = self._server.downpour_server_param.downpour_table_param.add()
         table.table_id = table_id
         table.table_class = "DownpourSparseTable"
         table.type = pslib.PS_SPARSE_TABLE
@@ -123,7 +123,7 @@ class DownpourServer(Server):
         Returns:
             return None 
         """
-        table = self.server_.downpour_server_param.downpour_table_param.add()
+        table = self._server.downpour_server_param.downpour_table_param.add()
         table.table_id = table_id
         table.table_class = "DownpourDenseTable"
         table.type = pslib.PS_DENSE_TABLE
@@ -140,7 +140,7 @@ class DownpourServer(Server):
         """
         Return downpour server program_desc
         """
-        return self.server_
+        return self._server
 
 
 class DownpourWorker(Worker):
@@ -155,7 +155,7 @@ class DownpourWorker(Worker):
 
     def __init__(self, window):
         self.window = window
-        self.worker_ = pslib.DownpourTrainerParameter()
+        self._worker = pslib.DownpourTrainerParameter()
 
     def add_sparse_table(self, table_id, learning_rate, slot_key_vars,
                          slot_value_vars):
@@ -187,7 +187,7 @@ class DownpourWorker(Worker):
         Returns:
             return None 
         """
-        table = self.worker_.dense_table.add()
+        table = self._worker.dense_table.add()
         table.table_id = table_id
         table.dense_variable_name.extend(
             filter(lambda x: x.find("embedding") == -1,
@@ -200,4 +200,4 @@ class DownpourWorker(Worker):
         """
         Return downpour worker program_desc
         """
-        return self.worker_
+        return self._worker
diff --git a/python/paddle/fluid/incubate/fleet/parameter_server/optimizer_factory.py b/python/paddle/fluid/incubate/fleet/parameter_server/optimizer_factory.py
index 94f79e77e72bfa2d0a09502722ef36d474b610b2..ba1f2c8f6ba43bcdb8d4240e33210370e5a454f6 100644
--- a/python/paddle/fluid/incubate/fleet/parameter_server/optimizer_factory.py
+++ b/python/paddle/fluid/incubate/fleet/parameter_server/optimizer_factory.py
@@ -24,9 +24,9 @@ from .node import DownpourWorker, DownpourServer
 
 class DistributedOptimizerImplBase(object):
     def __init__(self, optimizer):
-        self.optimizer_ = optimizer
-        self.learning_rate_ = optimizer._learning_rate
-        self.regularization_ = optimizer.regularization
+        self._optimizer = optimizer
+        self._learning_rate = optimizer._learning_rate
+        self._regularization = optimizer.regularization
 
     def minimize(self,
                  losses,
@@ -41,7 +41,7 @@ class DistributedAdam(DistributedOptimizerImplBase):
         # todo(guru4elephant): add more optimizers here as argument
         # todo(guru4elephant): make learning_rate as a variable
         super(DistributedAdam, self).__init__(optimizer)
-        self.window_ = 1
+        self._window = 1
         self.type = "downpour"
         self.data_norm_name = [
             ".batch_size", ".batch_square_sum", ".batch_sum",
@@ -79,9 +79,9 @@ class DistributedAdam(DistributedOptimizerImplBase):
         server = DownpourServer()
         worker = DownpourWorker(self.window_)
         sparse_table_index = 0
-        server.add_sparse_table(sparse_table_index, self.learning_rate_,
+        server.add_sparse_table(sparse_table_index, self._learning_rate,
                                 prefetch_slots, prefetch_slots_emb)
-        worker.add_sparse_table(sparse_table_index, self.learning_rate_,
+        worker.add_sparse_table(sparse_table_index, self._learning_rate,
                                 prefetch_slots, prefetch_slots_emb)
         dense_table_index = 1
         program_configs = {}
@@ -124,9 +124,9 @@ class DistributedAdam(DistributedOptimizerImplBase):
                         data_norm_grads.append(i[1])
                 if not is_data_norm_data:
                     grads.append(i[1])
-            server.add_dense_table(dense_table_index, self.learning_rate_,
+            server.add_dense_table(dense_table_index, self._learning_rate,
                                    params, grads)
-            worker.add_dense_table(dense_table_index, self.learning_rate_,
+            worker.add_dense_table(dense_table_index, self._learning_rate,
                                    params, grads)
             program_configs[program_id]["pull_dense"] = [dense_table_index]
             program_configs[program_id]["push_dense"] = [dense_table_index]
@@ -135,9 +135,9 @@ class DistributedAdam(DistributedOptimizerImplBase):
             if len(data_norm_params) != 0 and len(data_norm_grads) != 0:
                 dense_table_index += 1
                 server.add_data_norm_table(dense_table_index,
-                                           self.learning_rate_,
+                                           self._learning_rate,
                                            data_norm_params, data_norm_grads)
-                worker.add_dense_table(dense_table_index, self.learning_rate_,
+                worker.add_dense_table(dense_table_index, self._learning_rate,
                                        data_norm_params, data_norm_grads)
                 #program_config.pull_dense_table_id.extend([dense_table_index])
                 #program_config.push_dense_table_id.extend([dense_table_index])
diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py
index a5e513ed5e35d530dd07c49339995461da8454a1..f8f461853f34a09eb2317f6ac93ad385cca3609f 100644
--- a/python/paddle/fluid/layers/control_flow.py
+++ b/python/paddle/fluid/layers/control_flow.py
@@ -267,8 +267,44 @@ class StaticRNN(object):
     """
     StaticRNN class.
 
-    StaticRNN class is used to create a StaticRNN. The RNN will have its
-    own parameters like inputs, outputs, memories, status and length.
+    The StaticRNN can process a batch of sequence data. The length of each
+    sample sequence must be equal. The StaticRNN will have its own parameters
+    like inputs, outputs, memories. **Note that the first dimension of inputs
+    represents sequence length, and all the sequence length of inputs must be
+    the same. And the meaning of each axis of input and output are the same.**
+
+    Examples:
+        >>> import paddle.fluid as fluid
+        >>> import paddle.fluid.layers as layers
+        >>>
+        >>> vocab_size, hidden_size=10000, 200
+        >>> x = layers.data(name="x", shape=[-1, 1, 1], dtype='int64')
+        >>> x_emb = layers.embedding(
+        >>>         input=x,
+        >>>         size=[vocab_size, hidden_size],
+        >>>         dtype='float32',
+        >>>         is_sparse=False)
+        >>> x_emb = layers.transpose(x_emb, perm=[1, 0, 2])
+        >>>
+        >>> rnn = fluid.layers.StaticRNN()
+        >>> with rnn.step():
+        >>>    word = rnn.step_input(x_emb)
+        >>>    prev = rnn.memory(shape=[-1, hidden_size], batch_ref = word)
+        >>>    hidden = fluid.layers.fc(input=[word, prev], size=hidden_size, act='relu')
+        >>>    rnn.update_memory(prev, hidden)  # set prev to hidden
+        >>>    rnn.step_output(hidden)
+        >>>
+        >>> result = rnn()
+
+    The StaticRNN will unfold sequence into time steps. Users need to define
+    how to process each time step during the :code:`with` step.
+
+    The :code:`memory` is used as a staging data cross time step. The initial
+    value of memory can be a variable that is filled with a constant value or
+    a specified variable.
+
+    The StaticRNN can mark multiple variables as its output. Use `rnn()` to
+    get the output sequence.
     """
     BEFORE_RNN_BLOCK = 0
     IN_RNN_BLOCK = 1
@@ -284,6 +320,9 @@ class StaticRNN(object):
         self.seq_len = None
 
     def step(self):
+        """
+        The block for user to define operators in RNN.
+        """
         return BlockGuardWithCompletion(self)
 
     def _assert_in_rnn_block_(self, method):
@@ -298,13 +337,28 @@ class StaticRNN(object):
                init_batch_dim_idx=0,
                ref_batch_dim_idx=1):
         """
+        Create a memory variable for static rnn.
+
+        If the :code:`init` is not None, :code:`memory` will be initialized by
+        this Variable. If the :code:`init` is None, :code:`shape` and :code:`batch_ref`
+        must be set, and this function will initialize a :code:`init` Variable.
+
         Args:
-            init: boot memory, if not set, a shape, batch_ref must be provided
-            shape: shape of the boot memory
-            batch_ref: batch size reference variable
-            init_value: the init value of boot memory
-            init_batch_dim_idx: the index of batch size in init's dimension
-            ref_batch_dim_idx: the index of batch size in batch_ref's dimension
+            init(Variable|None): The initialized variable. If it is not set,
+                :code:`shape` and :code:`batch_ref` must be provided.
+                Default: None.
+            shape(list|tuple): The shape of the boot memory. NOTE the shape
+                does not contain batch_size. Default: None.
+            batch_ref(Variable|None): The batch size reference Variable.
+                Default: None.
+            init_value(float): the init value of boot memory. Default: 0.0.
+            init_batch_dim_idx(int): the batch_size axis of the
+                :code:`init` Variable. Default: 0.
+            ref_batch_dim_idx(int): the batch_size axis of the
+                :code:`batch_ref` Variable. Default: 1.
+
+        Returns:
+            The memory variable.
         """
         self._assert_in_rnn_block_('memory')
         if init is None:
@@ -343,6 +397,16 @@ class StaticRNN(object):
             return pre_mem
 
     def step_input(self, x):
+        """
+        Mark a sequence as a StaticRNN input.
+
+        Args:
+            x(Variable): The input sequence, the shape of x
+                should be [seq_len, ...].
+
+        Returns:
+            The current time step in the input sequence.
+        """
         self._assert_in_rnn_block_('step_input')
         if not isinstance(x, Variable):
             raise TypeError("step input takes a Variable")
@@ -357,6 +421,15 @@ class StaticRNN(object):
         return ipt
 
     def step_output(self, o):
+        """
+        Mark a sequence as a StaticRNN output.
+
+        Args:
+            o(Variable): The output sequence.
+
+        Returns:
+            None.
+        """
         self._assert_in_rnn_block_('step_output')
         if not isinstance(o, Variable):
             raise TypeError("step output takes a Variable")
@@ -376,10 +449,30 @@ class StaticRNN(object):
         self.outputs.append(out_var)
 
     def output(self, *outputs):
+        """
+        Mark the StaticRNN output variables.
+
+        Args:
+            outputs: The output Variables.
+
+        Returns:
+            None
+        """
         for each in outputs:
             self.step_output(each)
 
     def update_memory(self, mem, var):
+        """
+        Update the memory from ex_mem to new_mem. NOTE that the shape and data
+        type of :code:`ex_mem` and :code:`new_mem` must be same.
+
+        Args:
+            mem(Variable): the memory variable.
+            var(Variable): the plain variable generated in RNN block.
+
+        Returns:
+            None
+        """
         if not isinstance(mem, Variable) or not isinstance(var, Variable):
             raise TypeError("update memory should take variables")
         self.memories[mem.name].mem = var
@@ -419,6 +512,9 @@ class StaticRNN(object):
         for m in self.memories:
             local_inputs.add(m)
 
+        # NOTE(zcd): the params have two categories of variables.
+        #   - the variables that are the out of StaticRnn.
+        #   - the variables that are the parameters of some layers, for example, conv2d.
         params = list()
         for op in rnn_block.ops:
             assert isinstance(op, Operator)
@@ -435,17 +531,19 @@ class StaticRNN(object):
         inlinks = [parent_block.var(i.name) for i in self.inputs]
         outlinks = self.outputs
 
+        # NOTE(zcd): the states maybe empty in some case.
         boot_memories = []
         pre_memories = []
         memories = []
         for _, mem in six.iteritems(self.memories):
             boot_memories.append(mem.init)
             pre_memories.append(mem.pre_mem.name)
+            assert mem.mem is not None, "%s should be updated in every step." % (
+                mem.init.name)
             mem_var = rnn_block.var(mem.mem.name)
             assert isinstance(mem_var, Variable)
             new_mem = self.helper.create_variable_for_type_inference(
                 dtype=mem_var.dtype)
-
             rnn_block.append_op(
                 type='rnn_memory_helper',
                 inputs={'X': [mem_var]},
@@ -464,6 +562,7 @@ class StaticRNN(object):
             outputs={'outputs': outlinks,
                      'step_scopes': [step_scope]},
             attrs={
+                'has_states': len(pre_memories) > 0,
                 'ex_states': pre_memories,
                 'states': memories,
                 'sub_block': rnn_block
diff --git a/python/paddle/fluid/layers/learning_rate_scheduler.py b/python/paddle/fluid/layers/learning_rate_scheduler.py
index b7d1eeba80d93d549a019455087bb7cc1d2a1083..a67c8058f2c42713738420e81316452e15acb697 100644
--- a/python/paddle/fluid/layers/learning_rate_scheduler.py
+++ b/python/paddle/fluid/layers/learning_rate_scheduler.py
@@ -35,8 +35,8 @@ from ..dygraph import learning_rate_scheduler as imperate_lr
 
 __all__ = [
     'exponential_decay', 'natural_exp_decay', 'inverse_time_decay',
-    'polynomial_decay', 'piecewise_decay', 'noam_decay', 'append_LARS',
-    'cosine_decay', 'linear_lr_warmup'
+    'polynomial_decay', 'piecewise_decay', 'noam_decay', 'cosine_decay',
+    'linear_lr_warmup'
 ]
 
 
@@ -349,24 +349,26 @@ def cosine_decay(learning_rate, step_each_epoch, epochs):
     training progresses. By using this function, the learning rate will be decayed by
     following cosine decay strategy.
 
-    decayed_lr = learning_rate * 0.5 * (math.cos(epoch * math.pi / epochs) + 1)
+    .. math::
+
+	decayed\_lr = learning\_rate * 0.5 * (math.cos * (epoch * \\frac{math.pi}{epochs} ) + 1)
     
     Args:
         learning_rate(Variable|float): The initial learning rate.
         step_each_epoch(int): the number of steps in an epoch.
         epochs(int): the number of epochs.
 
-     Returns:
-        Variable: The decayed learning rate.
-
-     Examples:
+    Returns:
+	Variable: The decayed learning rate.
 
-    ..code-block:: python
+    Examples:
+	.. code-block:: python
 
-  	base_lr = 0.1
-	lr = fluid.layers.cosine_decay(
-	learning_rate = base_lr, step_each_epoch=10000, epochs=120)
+  	    base_lr = 0.1
+	    lr = fluid.layers.cosine_decay(
+	    learning_rate = base_lr, step_each_epoch=10000, epochs=120)
     """
+
     with default_main_program()._lr_schedule_guard():
         if imperative_base.enabled():
             decay = imperate_lr.CosineDecay(learning_rate, step_each_epoch,
@@ -381,50 +383,6 @@ def cosine_decay(learning_rate, step_each_epoch, epochs):
             return decayed_lr
 
 
-def append_LARS(params_grads, learning_rate, weight_decay):
-    """
-    Applies LARS (LAYER-WISE ADAPTIVE RATE SCALING) to learning rate for
-    each layer.
-
-    Args:
-        learning_rate: A learning rate Variable. This
-          is the global learning rate for LARS.
-        weight_decay: A Python `float` number.
-
-    Returns:
-        The decayed learning rate
-    Examples:
-        .. code-block:: python
-
-            learning_rate *= local_gw_ratio * sqrt(sumsq(param))
-                        / (sqrt(sumsq(gradient))+ weight_decay * sqrt(sumsq(param)))
-    """
-
-    assert not imperative_base.enabled(
-    ), "append_LARS is NOT supported in dygraph mode now"
-
-    def _balanced_weight(param_norm, grad_norm):
-        if weight_decay == 1.0:
-            return grad_norm + param_norm
-        else:
-            return grad_norm + weight_decay * param_norm
-
-    for param, grad in params_grads:
-        with param.block.program.optimized_guard(
-            [param, grad]), name_scope("optimizer"):
-            param_lr = param.optimize_attr['learning_rate']
-            param_norm = ops.sqrt(nn.reduce_sum(input=ops.square(param)))
-            grad_norm = ops.sqrt(nn.reduce_sum(input=ops.square(grad)))
-            if type(param_lr) == float and param_lr == 1.0:
-                decayed_lr = learning_rate * param_norm \
-                    / _balanced_weight(param_norm, grad_norm)
-            else:
-                decayed_lr = learning_rate * param_lr * param_norm \
-                    / _balanced_weight(param_norm, grad_norm)
-            # set back param local learning rate
-            param.optimize_attr['learning_rate'] = decayed_lr
-
-
 def linear_lr_warmup(learning_rate, warmup_steps, start_lr, end_lr):
     """
     Applies linear learning rate warmup before the normal learning rate
diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py
index e70ea9da38f7640fa09a39e7f8cbfeb91ddc9832..93e46eef16fb177169db679a8437d9a33ed38e99 100644
--- a/python/paddle/fluid/layers/nn.py
+++ b/python/paddle/fluid/layers/nn.py
@@ -73,6 +73,8 @@ __all__ = [
     'reduce_max',
     'reduce_min',
     'reduce_prod',
+    'reduce_all',
+    'reduce_any',
     'sequence_first_step',
     'sequence_last_step',
     'sequence_slice',
@@ -159,6 +161,7 @@ __all__ = [
     'sum',
     'slice',
     'shape',
+    'rank',
     'logical_and',
     'logical_or',
     'logical_xor',
@@ -4738,6 +4741,106 @@ def reduce_prod(input, dim=None, keep_dim=False, name=None):
     return out
 
 
+def reduce_all(input, dim=None, keep_dim=False, name=None):
+    """
+    Computes the ``logical and`` of tensor elements over the given dimension.
+
+    Args:
+        input (Variable): The input variable which is a Tensor or LoDTensor.
+        dim (list|int|None): The dimension along which the logical and is computed.
+            If :attr:`None`, compute the logical and over all elements of
+            :attr:`input` and return a Tensor variable with a single element,
+            otherwise must be in the range :math:`[-rank(input), rank(input))`.
+            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
+        keep_dim (bool): Whether to reserve the reduced dimension in the
+            output Tensor. The result tensor will have one fewer dimension
+            than the :attr:`input` unless :attr:`keep_dim` is true.
+        name(str|None): A name for this layer(optional). If set None, the layer
+                       will be named automatically.
+
+    Returns:
+        Variable: The reduced Tensor variable.
+
+    Examples:
+        .. code-block:: python
+        
+            # x is a bool Tensor variable with following elements:
+            #    [[True, False]
+            #     [True, True]]
+            # Each example is followed by the correspending output tensor.
+            fluid.layers.reduce_all(x)  # False 
+            fluid.layers.reduce_all(x, dim=0)  # [True, False]
+            fluid.layers.reduce_all(x, dim=-1)  # [False, True]
+            fluid.layers.reduce_all(x, dim=1,
+                                     keep_dim=True)  # [[False], [True]]
+
+    """
+    helper = LayerHelper('reduce_all', **locals())
+    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
+    if dim is not None and not isinstance(dim, list):
+        dim = [dim]
+    helper.append_op(
+        type='reduce_all',
+        inputs={'X': input},
+        outputs={'Out': out},
+        attrs={
+            'dim': dim if dim != None else [0],
+            'keep_dim': keep_dim,
+            'reduce_all': True if dim == None else False
+        })
+    return out
+
+
+def reduce_any(input, dim=None, keep_dim=False, name=None):
+    """
+    Computes the ``logical or`` of tensor elements over the given dimension.
+
+    Args:
+        input (Variable): The input variable which is a Tensor or LoDTensor.
+        dim (list|int|None): The dimension along which the logical or is computed.
+            If :attr:`None`, compute the logical or over all elements of
+            :attr:`input` and return a Tensor variable with a single element,
+            otherwise must be in the range :math:`[-rank(input), rank(input))`.
+            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
+        keep_dim (bool): Whether to reserve the reduced dimension in the
+            output Tensor. The result tensor will have one fewer dimension
+            than the :attr:`input` unless :attr:`keep_dim` is true.
+        name(str|None): A name for this layer(optional). If set None, the layer
+                       will be named automatically.
+
+    Returns:
+        Variable: The reduced Tensor variable.
+
+    Examples:
+        .. code-block:: python
+
+            # x is a bool Tensor variable with following elements:
+            #    [[True, False]
+            #     [False, False]]
+            # Each example is followed by the correspending output tensor.
+            fluid.layers.reduce_any(x)  # True
+            fluid.layers.reduce_any(x, dim=0)  # [True, False]
+            fluid.layers.reduce_any(x, dim=-1)  # [True, False]
+            fluid.layers.reduce_any(x, dim=1,
+                                     keep_dim=True)  # [[True], [False]]
+
+    """
+    helper = LayerHelper('reduce_any', **locals())
+    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
+    if dim is not None and not isinstance(dim, list):
+        dim = [dim]
+    helper.append_op(
+        type='reduce_any',
+        inputs={'X': input},
+        outputs={'Out': out},
+        attrs={
+            'dim': dim if dim != None else [0],
+            'keep_dim': keep_dim,
+            'reduce_all': True if dim == None else False
+        })
+    return out
+
+
 def split(input, num_or_sections, dim=-1, name=None):
     """
     Split the input tensor into multiple sub-tensors.
@@ -4819,7 +4922,7 @@ def l2_normalize(x, axis, epsilon=1e-12, name=None):
             the dimension to normalization is rank(X) + axis. -1 is the
             last dimension.
         epsilon(float): The epsilon value is used to avoid division by zero, \
-            the defalut value is 1e-10.
+            the defalut value is 1e-12.
         name(str|None): A name for this layer(optional). If set None, the layer \
             will be named automatically.
 
@@ -9237,6 +9340,32 @@ def shape(input):
     return out
 
 
+def rank(input):
+    """
+    **Rank Layer**
+
+    Returns the number of dimensions for a tensor, which is a 0-D int32 Tensor.
+
+    Args:
+        input (Variable): The input variable.
+
+    Returns:
+        Variable: The rank of the input variable.
+
+    Examples:
+        .. code-block:: python
+
+            input = layers.data(
+                name="input", shape=[3, 100, 100], dtype="float32")
+            rank = layers.rank(input) # 4
+    """
+
+    ndims = len(input.shape)
+    out = assign(np.array(ndims, 'int32'))
+
+    return out
+
+
 def _elementwise_op(helper):
     op_type = helper.layer_type
     x = helper.kwargs.get('x', None)
@@ -11002,7 +11131,7 @@ def pixel_shuffle(x, upscale_factor):
 
     Returns:
 
-        Out(Variable): the pixel shuffle result is a tensor variable with the same shape and the same type as the input.
+        Out(Variable): Reshaped tensor according to the new dimension.
 
     Raises:
 
diff --git a/python/paddle/fluid/layers/tensor.py b/python/paddle/fluid/layers/tensor.py
index 80450119f44e93aae4b483983484ea18be5b2035..03ebd41fa00c69bfce66d325e32fc9aeb25a2486 100644
--- a/python/paddle/fluid/layers/tensor.py
+++ b/python/paddle/fluid/layers/tensor.py
@@ -24,26 +24,11 @@ from .layer_function_generator import templatedoc
 import numpy
 
 __all__ = [
-    'create_tensor',
-    'create_parameter',
-    'create_global_var',
-    'cast',
-    'tensor_array_to_tensor',
-    'concat',
-    'sums',
-    'assign',
-    'fill_constant_batch_size_like',
-    'fill_constant',
-    'argmin',
-    'argmax',
-    'argsort',
-    'ones',
-    'zeros',
-    'reverse',
-    'has_inf',
-    'has_nan',
-    'isfinite',
-    'range',
+    'create_tensor', 'create_parameter', 'create_global_var', 'cast',
+    'tensor_array_to_tensor', 'concat', 'sums', 'assign',
+    'fill_constant_batch_size_like', 'fill_constant', 'argmin', 'argmax',
+    'argsort', 'ones', 'zeros', 'reverse', 'has_inf', 'has_nan', 'isfinite',
+    'range', 'linspace'
 ]
 
 
@@ -826,3 +811,45 @@ def range(start, end, step, dtype):
                 'Step': step},
         outputs={'Out': [out]})
     return out
+
+
+def linspace(start, stop, num, dtype):
+    """
+    Return fixed number of evenly spaced values within a given interval.
+
+    First entry is start, and last entry is stop. In the case when Num is 1, only Start is returned. Like linspace function of numpy.
+
+    Args:
+        start(float|Variable): First entry in the sequence. It is a float scalar, or a tensor of shape [1] with type 'float32'|'float64'.
+        stop(float|Variable): Last entry in the sequence. It is a float scalar, or a tensor of shape [1] with type 'float32'|'float64'.
+        num(int|Variable): Number of entry in the sequence. It is an int scalar, or a tensor of shape [1] with type int32.
+        dtype(string): 'float32'|'float64', the data type of the output tensor.
+
+    Returns:
+        Variable: The tensor variable storing a 1-D tensor. 
+
+    Examples:
+        .. code-block:: python
+
+             data = fluid.layers.linspace(0, 10, 5, 'float32') # [0.0,  2.5,  5.0,  7.5, 10.0]
+             data = fluid.layers.linspace(0, 10, 1, 'float32') # [0.0]
+
+    """
+    helper = LayerHelper("linspace", **locals())
+
+    if not isinstance(start, Variable):
+        start = fill_constant([1], dtype, start)
+    if not isinstance(stop, Variable):
+        stop = fill_constant([1], dtype, stop)
+    if not isinstance(num, Variable):
+        num = fill_constant([1], 'int32', num)
+
+    out = helper.create_variable_for_type_inference(dtype=start.dtype)
+
+    helper.append_op(
+        type='linspace',
+        inputs={'Start': start,
+                'Stop': stop,
+                'Num': num},
+        outputs={'Out': [out]})
+    return out
diff --git a/python/paddle/fluid/tests/book/CMakeLists.txt b/python/paddle/fluid/tests/book/CMakeLists.txt
index ee734f3c782adb5196a03aca5718377009a5b4e7..999a765b6dc32323a24f9069f11134360dbadcb8 100644
--- a/python/paddle/fluid/tests/book/CMakeLists.txt
+++ b/python/paddle/fluid/tests/book/CMakeLists.txt
@@ -6,4 +6,6 @@ foreach(src ${TEST_OPS})
     py_test(${src} SRCS ${src}.py)
 endforeach()
 
-add_subdirectory(high-level-api)
+if(WITH_HIGH_LEVEL_API_TEST)
+  add_subdirectory(high-level-api)
+endif()
diff --git a/python/paddle/fluid/tests/book/high-level-api/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/CMakeLists.txt
index efa5ee2d06af3d31e7d84122dd7eea37d6dcf3a3..c034709fbdc2aa315ca995a42c278b261e6283a4 100644
--- a/python/paddle/fluid/tests/book/high-level-api/CMakeLists.txt
+++ b/python/paddle/fluid/tests/book/high-level-api/CMakeLists.txt
@@ -1,16 +1,28 @@
-file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
+file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*_new_api.py")
 string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
 
-# default test
-foreach(src ${TEST_OPS})
-    py_test(${src} SRCS ${src}.py)
-endforeach()
+# This test is buggy
+# py_test(test_understand_sentiment_dynamic_rnn SRCS
+# 	test_understand_sentiment_dynamic_rnn.py SERIAL)
+LIST(REMOVE_ITEM TEST_OPS test_understand_sentiment_dynamic_rnn_new_api)
 
-add_subdirectory(fit_a_line)
-add_subdirectory(recognize_digits)
-add_subdirectory(image_classification)
-add_subdirectory(understand_sentiment)
-add_subdirectory(label_semantic_roles)
-add_subdirectory(word2vec)
-add_subdirectory(recommender_system)
-add_subdirectory(machine_translation)
+if(NOT APPLE)
+    # default test
+    foreach(src ${TEST_OPS})
+        py_test(${src} SRCS ${src}.py)
+    endforeach()
+else()
+    foreach(src ${TEST_OPS})
+        if(${src} STREQUAL "test_image_classification_vgg_new_api")
+            message(WARNING "These tests has been disabled in OSX for random fail: \n" ${src})
+        elseif(${src} STREQUAL "test_image_classification_resnet_new_api")
+            message(WARNING "These tests has been disabled in OSX for random fail: \n" ${src})
+        elseif(${src} STREQUAL "test_recognize_digits_conv_new_api")
+            message(WARNING "These tests has been disabled in OSX for random fail: \n" ${src})
+        elseif(${src} STREQUAL "test_recognize_digits_mlp_new_api")
+            message(WARNING "These tests has been disabled in OSX for random fail: \n" ${src})
+        elseif()
+            py_test(${src} SRCS ${src}.py)
+        endif()
+    endforeach()
+endif()
diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py b/python/paddle/fluid/tests/book/high-level-api/cifar10_small_test_set.py
similarity index 100%
rename from python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py
rename to python/paddle/fluid/tests/book/high-level-api/cifar10_small_test_set.py
diff --git a/python/paddle/fluid/tests/book/high-level-api/fit_a_line/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/fit_a_line/CMakeLists.txt
deleted file mode 100644
index 673c965b662a022739f8d489c331f4de9455a926..0000000000000000000000000000000000000000
--- a/python/paddle/fluid/tests/book/high-level-api/fit_a_line/CMakeLists.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
-string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
-
-# default test
-foreach(src ${TEST_OPS})
-    py_test(${src} SRCS ${src}.py)
-endforeach()
diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/image_classification/CMakeLists.txt
deleted file mode 100644
index 91c1d17eb5391ea37a41a886594cc71c6e6c56bd..0000000000000000000000000000000000000000
--- a/python/paddle/fluid/tests/book/high-level-api/image_classification/CMakeLists.txt
+++ /dev/null
@@ -1,19 +0,0 @@
-file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
-string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
-
-if(NOT APPLE)
-    # default test
-    foreach(src ${TEST_OPS})
-        py_test(${src} SRCS ${src}.py)
-    endforeach()
-else()
-    foreach(src ${TEST_OPS})
-        if(${src} STREQUAL "test_image_classification_vgg")
-            message(WARNING "These tests has been disabled in OSX for random fail: \n" ${src})
-        elseif(${src} STREQUAL "test_image_classification_resnet")
-            message(WARNING "These tests has been disabled in OSX for random fail: \n" ${src})
-        elseif()
-            py_test(${src} SRCS ${src}.py)
-        endif()
-    endforeach()
-endif()
diff --git a/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/CMakeLists.txt
deleted file mode 100644
index 673c965b662a022739f8d489c331f4de9455a926..0000000000000000000000000000000000000000
--- a/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/CMakeLists.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
-string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
-
-# default test
-foreach(src ${TEST_OPS})
-    py_test(${src} SRCS ${src}.py)
-endforeach()
diff --git a/python/paddle/fluid/tests/book/high-level-api/machine_translation/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/machine_translation/CMakeLists.txt
deleted file mode 100644
index 673c965b662a022739f8d489c331f4de9455a926..0000000000000000000000000000000000000000
--- a/python/paddle/fluid/tests/book/high-level-api/machine_translation/CMakeLists.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
-string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
-
-# default test
-foreach(src ${TEST_OPS})
-    py_test(${src} SRCS ${src}.py)
-endforeach()
diff --git a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/CMakeLists.txt
deleted file mode 100644
index f9c6d60540fcb6f8a73fdc4e68471448e16cbdc2..0000000000000000000000000000000000000000
--- a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/CMakeLists.txt
+++ /dev/null
@@ -1,19 +0,0 @@
-file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
-string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
-
-# default test
-if(NOT APPLE)
-    foreach(src ${TEST_OPS})
-        py_test(${src} SRCS ${src}.py)
-    endforeach()
-else()
-    foreach(src ${TEST_OPS})
-        if(${src} STREQUAL "test_recognize_digits_conv")
-            message(WARNING "These tests has been disabled in OSX for random fail: \n" ${src})
-        elseif(${src} STREQUAL "test_recognize_digits_mlp")
-            message(WARNING "These tests has been disabled in OSX for random fail: \n" ${src})
-        else()
-            py_test(${src} SRCS ${src}.py)
-        endif()
-    endforeach()
-endif()
diff --git a/python/paddle/fluid/tests/book/high-level-api/recommender_system/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/recommender_system/CMakeLists.txt
deleted file mode 100644
index 673c965b662a022739f8d489c331f4de9455a926..0000000000000000000000000000000000000000
--- a/python/paddle/fluid/tests/book/high-level-api/recommender_system/CMakeLists.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
-string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
-
-# default test
-foreach(src ${TEST_OPS})
-    py_test(${src} SRCS ${src}.py)
-endforeach()
diff --git a/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py b/python/paddle/fluid/tests/book/high-level-api/test_fit_a_line_new_api.py
similarity index 100%
rename from python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py
rename to python/paddle/fluid/tests/book/high-level-api/test_fit_a_line_new_api.py
diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py b/python/paddle/fluid/tests/book/high-level-api/test_image_classification_resnet_new_api.py
similarity index 100%
rename from python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py
rename to python/paddle/fluid/tests/book/high-level-api/test_image_classification_resnet_new_api.py
diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py b/python/paddle/fluid/tests/book/high-level-api/test_image_classification_vgg_new_api.py
similarity index 100%
rename from python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py
rename to python/paddle/fluid/tests/book/high-level-api/test_image_classification_vgg_new_api.py
diff --git a/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/test_label_semantic_roles_newapi.py b/python/paddle/fluid/tests/book/high-level-api/test_label_semantic_roles_new_api.py
similarity index 100%
rename from python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/test_label_semantic_roles_newapi.py
rename to python/paddle/fluid/tests/book/high-level-api/test_label_semantic_roles_new_api.py
diff --git a/python/paddle/fluid/tests/book/high-level-api/machine_translation/test_machine_translation.py b/python/paddle/fluid/tests/book/high-level-api/test_machine_translation_new_api.py
similarity index 100%
rename from python/paddle/fluid/tests/book/high-level-api/machine_translation/test_machine_translation.py
rename to python/paddle/fluid/tests/book/high-level-api/test_machine_translation_new_api.py
diff --git a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py b/python/paddle/fluid/tests/book/high-level-api/test_recognize_digits_conv_new_api.py
similarity index 100%
rename from python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py
rename to python/paddle/fluid/tests/book/high-level-api/test_recognize_digits_conv_new_api.py
diff --git a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py b/python/paddle/fluid/tests/book/high-level-api/test_recognize_digits_mlp_new_api.py
similarity index 100%
rename from python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py
rename to python/paddle/fluid/tests/book/high-level-api/test_recognize_digits_mlp_new_api.py
diff --git a/python/paddle/fluid/tests/book/high-level-api/recommender_system/test_recommender_system_newapi.py b/python/paddle/fluid/tests/book/high-level-api/test_recommender_system_new_api.py
similarity index 100%
rename from python/paddle/fluid/tests/book/high-level-api/recommender_system/test_recommender_system_newapi.py
rename to python/paddle/fluid/tests/book/high-level-api/test_recommender_system_new_api.py
diff --git a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_conv.py b/python/paddle/fluid/tests/book/high-level-api/test_understand_sentiment_conv_new_api.py
similarity index 100%
rename from python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_conv.py
rename to python/paddle/fluid/tests/book/high-level-api/test_understand_sentiment_conv_new_api.py
diff --git a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_dynamic_rnn.py b/python/paddle/fluid/tests/book/high-level-api/test_understand_sentiment_dynamic_rnn_new_api.py
similarity index 100%
rename from python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_dynamic_rnn.py
rename to python/paddle/fluid/tests/book/high-level-api/test_understand_sentiment_dynamic_rnn_new_api.py
diff --git a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py b/python/paddle/fluid/tests/book/high-level-api/test_understand_sentiment_stacked_lstm_new_api.py
similarity index 100%
rename from python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py
rename to python/paddle/fluid/tests/book/high-level-api/test_understand_sentiment_stacked_lstm_new_api.py
diff --git a/python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py b/python/paddle/fluid/tests/book/high-level-api/test_word2vec_new_api.py
similarity index 100%
rename from python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py
rename to python/paddle/fluid/tests/book/high-level-api/test_word2vec_new_api.py
diff --git a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/CMakeLists.txt
deleted file mode 100644
index d71147a85e77ea6dc5b6391aa169abd9b02a0aa1..0000000000000000000000000000000000000000
--- a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/CMakeLists.txt
+++ /dev/null
@@ -1,12 +0,0 @@
-file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
-string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
-
-# This test is buggy
-# py_test(test_understand_sentiment_dynamic_rnn SRCS
-# 	test_understand_sentiment_dynamic_rnn.py SERIAL)
-LIST(REMOVE_ITEM TEST_OPS test_understand_sentiment_dynamic_rnn)
-
-# default test
-foreach(src ${TEST_OPS})
-    py_test(${src} SRCS ${src}.py)
-endforeach()
diff --git a/python/paddle/fluid/tests/book/high-level-api/word2vec/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/word2vec/CMakeLists.txt
deleted file mode 100644
index 673c965b662a022739f8d489c331f4de9455a926..0000000000000000000000000000000000000000
--- a/python/paddle/fluid/tests/book/high-level-api/word2vec/CMakeLists.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
-string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
-
-# default test
-foreach(src ${TEST_OPS})
-    py_test(${src} SRCS ${src}.py)
-endforeach()
diff --git a/python/paddle/fluid/tests/unittests/CMakeLists.txt b/python/paddle/fluid/tests/unittests/CMakeLists.txt
index cbe9afce035ea9918a41fafe3c2d4a3eb3f4dcb0..65045a4ab20338817908e481e5f990706cc88e71 100644
--- a/python/paddle/fluid/tests/unittests/CMakeLists.txt
+++ b/python/paddle/fluid/tests/unittests/CMakeLists.txt
@@ -25,7 +25,6 @@ endif()
 list(REMOVE_ITEM TEST_OPS test_seq_concat_op) # FIXME(helin): https://github.com/PaddlePaddle/Paddle/issues/8290
 list(REMOVE_ITEM TEST_OPS test_modified_huber_loss_op) # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/5184
 list(REMOVE_ITEM TEST_OPS test_lstm_unit_op) # # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/5185
-list(REMOVE_ITEM TEST_OPS test_recurrent_op) # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/6152
 list(REMOVE_ITEM TEST_OPS test_cond_op) # FIXME(qijun): https://github.com/PaddlePaddle/Paddle/issues/5101#issuecomment-339814957
 
 list(REMOVE_ITEM TEST_OPS op_test) # op_test is a helper python file, not a test
@@ -74,7 +73,6 @@ list(REMOVE_ITEM TEST_OPS test_dgc_op)
 list(REMOVE_ITEM TEST_OPS test_dist_se_resnext_nccl)
 list(REMOVE_ITEM TEST_OPS test_dist_transformer)
 list(REMOVE_ITEM TEST_OPS test_parallel_executor_transformer)
-list(REMOVE_ITEM TEST_OPS test_image_classification_resnet)
 list(REMOVE_ITEM TEST_OPS test_bilinear_interp_op)
 list(REMOVE_ITEM TEST_OPS test_nearest_interp_op)
 list(REMOVE_ITEM TEST_OPS test_imperative_resnet)
@@ -125,10 +123,6 @@ if(NOT WIN32)
     py_test_modules(test_ir_memory_optimize_transformer MODULES test_ir_memory_optimize_transformer SERIAL)
 endif()
 
-if(NOT APPLE)
-    py_test_modules(test_image_classification_resnet MODULES test_image_classification_resnet SERIAL)
-endif()
-
 if(CMAKE_BUILD_TYPE STREQUAL "Debug")
     # change the timeout from 600 to 2200, because in debug mode, this test need more time.
     set_tests_properties(test_parallel_executor_seresnext PROPERTIES TIMEOUT 2200)
diff --git a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py
index 18ed02a72275437fa6106e57c0383e17647d9700..723aafb171271ed248c93665a21089029a30a836 100644
--- a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py
+++ b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py
@@ -29,7 +29,8 @@ __all__ = ['TestParallelExecutorBase']
 
 
 class TestParallelExecutorBase(unittest.TestCase):
-    def check_network_convergence(self,
+    @classmethod
+    def check_network_convergence(cls,
                                   method,
                                   use_cuda=True,
                                   memory_opt=True,
diff --git a/python/paddle/fluid/tests/unittests/test_dyn_rnn.py b/python/paddle/fluid/tests/unittests/test_dyn_rnn.py
index 48fb93ec529bee32b9652a89ba7da3dc77f7853a..4b0195d307dc83f77ff04e89544d7bc751b8c011 100644
--- a/python/paddle/fluid/tests/unittests/test_dyn_rnn.py
+++ b/python/paddle/fluid/tests/unittests/test_dyn_rnn.py
@@ -24,15 +24,15 @@ from paddle.fluid.layers.control_flow import max_sequence_len
 from paddle.fluid.layers.control_flow import lod_tensor_to_array
 from paddle.fluid.layers.control_flow import array_to_lod_tensor
 from paddle.fluid.layers.control_flow import shrink_memory
+from fake_reader import fake_imdb_reader
 
 
 class TestDynRNN(unittest.TestCase):
     def setUp(self):
-        self.word_dict = paddle.dataset.imdb.word_dict()
+        self.word_dict_len = 5147
         self.BATCH_SIZE = 2
-        self.train_data = paddle.batch(
-            paddle.dataset.imdb.train(self.word_dict),
-            batch_size=self.BATCH_SIZE)
+        reader = fake_imdb_reader(self.word_dict_len, self.BATCH_SIZE * 100)
+        self.train_data = paddle.batch(reader, batch_size=self.BATCH_SIZE)
 
     def test_plain_while_op(self):
         main_program = fluid.Program()
@@ -42,7 +42,7 @@ class TestDynRNN(unittest.TestCase):
             sentence = fluid.layers.data(
                 name='word', shape=[1], dtype='int64', lod_level=1)
             sent_emb = fluid.layers.embedding(
-                input=sentence, size=[len(self.word_dict), 32], dtype='float32')
+                input=sentence, size=[self.word_dict_len, 32], dtype='float32')
 
             label = fluid.layers.data(name='label', shape=[1], dtype='float32')
 
@@ -109,7 +109,7 @@ class TestDynRNN(unittest.TestCase):
             sentence = fluid.layers.data(
                 name='word', shape=[1], dtype='int64', lod_level=1)
             sent_emb = fluid.layers.embedding(
-                input=sentence, size=[len(self.word_dict), 32], dtype='float32')
+                input=sentence, size=[self.word_dict_len, 32], dtype='float32')
 
             rnn = fluid.layers.DynamicRNN()
 
diff --git a/python/paddle/fluid/tests/unittests/test_layers.py b/python/paddle/fluid/tests/unittests/test_layers.py
index 38d0533a7ec820241c6a08f2180a7426984068f2..91f8bc5fd0a510dcc05cb7ba2397cad52be16af5 100644
--- a/python/paddle/fluid/tests/unittests/test_layers.py
+++ b/python/paddle/fluid/tests/unittests/test_layers.py
@@ -1759,10 +1759,20 @@ class TestBook(LayerTest):
     def test_lod_reset(self):
         # TODO(minqiyang): dygraph do not support lod now
         with self.static_graph():
+            # case 1
             x = layers.data(name='x', shape=[10], dtype='float32')
             y = layers.data(
                 name='y', shape=[10, 20], dtype='float32', lod_level=2)
-            return (layers.lod_reset(x=x, y=y))
+            z = layers.lod_reset(x=x, y=y)
+            self.assertTrue(z.lod_level == 2)
+            # case 2
+            lod_tensor_in = layers.data(name='lod_in', shape=[1], dtype='int64')
+            z = layers.lod_reset(x=x, y=lod_tensor_in)
+            self.assertTrue(z.lod_level == 1)
+            # case 3
+            z = layers.lod_reset(x=x, target_lod=[1, 2, 3])
+            self.assertTrue(z.lod_level == 1)
+            return z
 
     def test_affine_grid(self):
         with self.static_graph():
@@ -1925,6 +1935,13 @@ class TestBook(LayerTest):
             out = layers.flatten(x, axis=1, name="flatten")
             return (out)
 
+    def test_linspace(self):
+        program = Program()
+        with program_guard(program):
+            out = layers.linspace(20, 10, 5, 'float64')
+            self.assertIsNotNone(out)
+        print(str(program))
+
 
 if __name__ == '__main__':
     unittest.main()
diff --git a/python/paddle/fluid/tests/unittests/test_linspace.py b/python/paddle/fluid/tests/unittests/test_linspace.py
new file mode 100644
index 0000000000000000000000000000000000000000..eeecf178320327cc251f32bfe46c1622200339f4
--- /dev/null
+++ b/python/paddle/fluid/tests/unittests/test_linspace.py
@@ -0,0 +1,71 @@
+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestLinspaceOpCommonCase(OpTest):
+    def setUp(self):
+        self.op_type = "linspace"
+        dtype = 'float32'
+        self.inputs = {
+            'Start': np.array([0]).astype(dtype),
+            'Stop': np.array([10]).astype(dtype),
+            'Num': np.array([11]).astype('int32')
+        }
+
+        self.outputs = {'Out': np.arange(0, 11).astype(dtype)}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestLinspaceOpReverseCase(OpTest):
+    def setUp(self):
+        self.op_type = "linspace"
+        dtype = 'float32'
+        self.inputs = {
+            'Start': np.array([10]).astype(dtype),
+            'Stop': np.array([0]).astype(dtype),
+            'Num': np.array([11]).astype('int32')
+        }
+
+        self.outputs = {'Out': np.arange(10, -1, -1).astype(dtype)}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestLinspaceOpNumOneCase(OpTest):
+    def setUp(self):
+        self.op_type = "linspace"
+        dtype = 'float32'
+        self.inputs = {
+            'Start': np.array([10]).astype(dtype),
+            'Stop': np.array([0]).astype(dtype),
+            'Num': np.array([1]).astype('int32')
+        }
+
+        self.outputs = {'Out': np.array(10, dtype=dtype)}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py
index 12d854fb54ac30ff2eeed97c16a78198d92387fd..92a5c58c11773e97ca0bb5ff2c21cbc8df612d58 100644
--- a/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py
+++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py
@@ -29,7 +29,7 @@ import unittest
 import math
 import numpy as np
 from functools import partial
-
+os.environ['CPU_NUM'] = str(4)
 # FIXME(zcd): If the neural net has dropout_op, the output of ParallelExecutor
 # and Executor is different. Because, for ParallelExecutor, the dropout_op of
 # the neural net will be copied N copies(N is the number of device). This will
@@ -113,7 +113,6 @@ def bottleneck_block(input, num_filters, stride, cardinality, reduction_ratio):
     return fluid.layers.elementwise_add(x=short, y=scale, act='relu')
 
 
-batch_size = 12
 img_shape = [3, 224, 224]
 
 
@@ -181,43 +180,84 @@ def optimizer(learning_rate=0.01):
     return optimizer
 
 
+def _batch_size():
+    return 12
+
+
+def _iter(use_cuda):
+    if use_cuda:
+        return 10
+    return 2
+
+
+gpu_img, gpu_label = init_data(
+    batch_size=_batch_size(), img_shape=img_shape, label_range=999)
+cpu_img, cpu_label = init_data(
+    batch_size=_batch_size(), img_shape=img_shape, label_range=999)
+feed_dict_gpu = {"image": gpu_img, "label": gpu_label}
+feed_dict_cpu = {"image": cpu_img, "label": cpu_label}
+model = SE_ResNeXt50Small
+
+
+def _feed_dict(use_cuda):
+    if use_cuda:
+        return feed_dict_gpu
+    return feed_dict_cpu
+
+
+def _get_result_of_origin_model(use_cuda):
+    global remove_bn
+    global remove_dropout
+    remove_bn = True
+    remove_dropout = True
+    first_loss, last_loss = TestParallelExecutorBase.check_network_convergence(
+        model,
+        feed_dict=_feed_dict(use_cuda),
+        iter=_iter(use_cuda),
+        batch_size=_batch_size(),
+        use_cuda=use_cuda,
+        use_reduce=False,
+        optimizer=optimizer)
+
+    return first_loss, last_loss
+
+
+origin_cpu_first_loss, origin_cpu_last_loss = _get_result_of_origin_model(False)
+if core.is_compiled_with_cuda():
+    origin_gpu_first_loss, origin_gpu_last_loss = _get_result_of_origin_model(
+        True)
+
+
+def _get_origin_result(use_cuda):
+    if use_cuda:
+        assert core.is_compiled_with_cuda(), "Doesn't compiled with CUDA."
+        return origin_gpu_first_loss, origin_gpu_last_loss
+    return origin_cpu_first_loss, origin_cpu_last_loss
+
+
 class TestResnet(TestParallelExecutorBase):
-    @classmethod
-    def setUpClass(cls):
-        os.environ['CPU_NUM'] = str(4)
-        global remove_dropout
-        global remove_bn
-        remove_dropout = False
-        remove_bn = False
-
-    def _compare_reduce_and_allreduce(self,
-                                      model,
-                                      use_cuda,
-                                      iter=20,
-                                      delta2=1e-5):
+    def _compare_reduce_and_allreduce(self, use_cuda, delta2=1e-5):
         if use_cuda and not core.is_compiled_with_cuda():
             return
 
         global remove_bn
+        global remove_dropout
         remove_bn = True
+        remove_dropout = True
 
-        img, label = init_data(
-            batch_size=batch_size, img_shape=img_shape, label_range=999)
         all_reduce_first_loss, all_reduce_last_loss = self.check_network_convergence(
             model,
-            feed_dict={"image": img,
-                       "label": label},
-            iter=iter,
-            batch_size=batch_size,
+            feed_dict=_feed_dict(use_cuda),
+            iter=_iter(use_cuda),
+            batch_size=_batch_size(),
             use_cuda=use_cuda,
             use_reduce=False,
             optimizer=optimizer)
         reduce_first_loss, reduce_last_loss = self.check_network_convergence(
             model,
-            feed_dict={"image": img,
-                       "label": label},
-            iter=iter,
-            batch_size=batch_size,
+            feed_dict=_feed_dict(use_cuda),
+            iter=_iter(use_cuda),
+            batch_size=_batch_size(),
             use_cuda=use_cuda,
             use_reduce=True,
             optimizer=optimizer)
@@ -232,10 +272,9 @@ class TestResnet(TestParallelExecutorBase):
 
         all_reduce_first_loss_seq, all_reduce_last_loss_seq = self.check_network_convergence(
             model,
-            feed_dict={"image": img,
-                       "label": label},
-            iter=iter,
-            batch_size=batch_size,
+            feed_dict=_feed_dict(use_cuda),
+            iter=_iter(use_cuda),
+            batch_size=_batch_size(),
             use_cuda=use_cuda,
             use_reduce=False,
             optimizer=optimizer,
@@ -243,10 +282,9 @@ class TestResnet(TestParallelExecutorBase):
 
         reduce_first_loss_seq, reduce_last_loss_seq = self.check_network_convergence(
             model,
-            feed_dict={"image": img,
-                       "label": label},
-            iter=iter,
-            batch_size=batch_size,
+            feed_dict=_feed_dict(use_cuda),
+            iter=_iter(use_cuda),
+            batch_size=_batch_size(),
             use_cuda=use_cuda,
             use_reduce=True,
             optimizer=optimizer,
@@ -267,37 +305,28 @@ class TestResnet(TestParallelExecutorBase):
         for loss in zip(all_reduce_last_loss_seq, reduce_last_loss_seq):
             self.assertAlmostEquals(loss[0], loss[1], delta=delta2)
 
-    def _check_resnet_convergence(self,
-                                  model,
-                                  check_func_1,
-                                  check_func_2,
-                                  use_cuda,
-                                  iter=20,
-                                  delta2=1e-5,
-                                  compare_seperately=True):
+    def _compare_result_with_origin_model(self,
+                                          get_origin_result,
+                                          check_func_2,
+                                          use_cuda,
+                                          delta2=1e-5,
+                                          compare_seperately=True,
+                                          rm_drop_out=False,
+                                          rm_bn=False):
         if use_cuda and not core.is_compiled_with_cuda():
             return
 
-        global remove_dropout
         global remove_bn
-        remove_dropout = True
-        remove_bn = True
+        global remove_dropout
+        remove_bn = rm_bn or use_cuda
+        remove_dropout = rm_drop_out
 
-        img, label = init_data(
-            batch_size=batch_size, img_shape=img_shape, label_range=999)
-        func_1_first_loss, func_1_last_loss = check_func_1(
-            model,
-            feed_dict={"image": img,
-                       "label": label},
-            iter=iter,
-            batch_size=batch_size,
-            use_cuda=use_cuda)
+        func_1_first_loss, func_1_last_loss = get_origin_result(use_cuda)
         func_2_first_loss, func_2_last_loss = check_func_2(
             model,
-            feed_dict={"image": img,
-                       "label": label},
-            iter=iter,
-            batch_size=batch_size,
+            feed_dict=_feed_dict(use_cuda),
+            iter=_iter(use_cuda),
+            batch_size=_batch_size(),
             use_cuda=use_cuda)
 
         if compare_seperately:
@@ -311,97 +340,55 @@ class TestResnet(TestParallelExecutorBase):
             self.assertAlmostEquals(
                 np.mean(func_1_last_loss), func_2_last_loss[0], delta=delta2)
 
-    def _compare_with_fused_all_reduce(self,
-                                       model,
-                                       use_cuda,
-                                       iter=20,
-                                       delta2=1e-5):
-        if use_cuda and not core.is_compiled_with_cuda():
-            return
-
-        global remove_bn
-        remove_bn = True
-
-        img, label = init_data(
-            batch_size=batch_size, img_shape=img_shape, label_range=999)
-        all_reduce_first_loss, all_reduce_last_loss = self.check_network_convergence(
-            model,
-            feed_dict={"image": img,
-                       "label": label},
-            iter=iter,
-            batch_size=batch_size,
-            use_cuda=use_cuda,
-            fuse_all_reduce_ops=False,
-            optimizer=optimizer)
-        reduce_first_loss, reduce_last_loss = self.check_network_convergence(
-            model,
-            feed_dict={"image": img,
-                       "label": label},
-            iter=iter,
-            batch_size=batch_size,
-            use_cuda=use_cuda,
-            fuse_all_reduce_ops=True,
-            optimizer=optimizer)
-
-        for loss in zip(all_reduce_first_loss, reduce_first_loss):
-            self.assertAlmostEquals(loss[0], loss[1], delta=1e-5)
-        for loss in zip(all_reduce_last_loss, reduce_last_loss):
-            self.assertAlmostEquals(loss[0], loss[1], delta=delta2)
-
     def test_seresnext_with_reduce(self):
-        self._compare_reduce_and_allreduce(
-            model=SE_ResNeXt50Small, use_cuda=True, delta2=1e-2)
-        self._compare_reduce_and_allreduce(
-            model=SE_ResNeXt50Small, use_cuda=False, iter=5)
-
-    def test_seresnext_with_fused_all_reduce(self):
-        self._compare_with_fused_all_reduce(
-            model=SE_ResNeXt50Small, use_cuda=True, delta2=1e-3)
-        self._compare_with_fused_all_reduce(
-            model=SE_ResNeXt50Small, use_cuda=False, iter=2, delta2=1e-3)
+        self._compare_reduce_and_allreduce(use_cuda=False, delta2=1e-3)
+        self._compare_reduce_and_allreduce(use_cuda=True, delta2=1e-2)
 
     def test_seresnext_with_learning_rate_decay(self):
-        check_func_1 = partial(
-            self.check_network_convergence,
-            optimizer=optimizer,
-            use_parallel_executor=True)
+        # NOTE(zcd): This test is compare the result of use parallel_executor and executor,
+        # and the result of drop_out op and batch_norm op in this two executor
+        # have diff, so the two ops should be removed from the model.
+        check_func_1 = _get_origin_result
         check_func_2 = partial(
             self.check_network_convergence,
             optimizer=optimizer,
             use_parallel_executor=False)
-        self._check_resnet_convergence(
-            SE_ResNeXt50Small,
-            check_func_1,
-            check_func_2,
-            use_cuda=True,
-            compare_seperately=False)
-        self._check_resnet_convergence(
-            SE_ResNeXt50Small,
+        self._compare_result_with_origin_model(
             check_func_1,
             check_func_2,
             use_cuda=False,
+            rm_drop_out=True,
+            rm_bn=True,
             compare_seperately=False,
-            iter=2,
             delta2=1e-3)
+        self._compare_result_with_origin_model(
+            check_func_1,
+            check_func_2,
+            use_cuda=True,
+            rm_drop_out=True,
+            rm_bn=True,
+            compare_seperately=False)
 
-    def test_seresnext_with_fused_optimizer_ops(self):
-        check_func_1 = partial(
-            self.check_network_convergence, fuse_all_optimizer_ops=False)
+    def test_seresnext_with_fused_all_reduce(self):
+        # NOTE(zcd): In order to make the program faster,
+        # this unit test remove drop_out and batch_norm.
+        check_func_1 = _get_origin_result
         check_func_2 = partial(
-            self.check_network_convergence, fuse_all_optimizer_ops=True)
-        # TODO(zcd): this test failed random, I will fix it in next PR.
-        # self._check_resnet_convergence(
-        #     SE_ResNeXt50Small,
-        #     check_func_1,
-        #     check_func_2,
-        #     use_cuda=True,
-        #     delta2=1e-3)
-        self._check_resnet_convergence(
-            SE_ResNeXt50Small,
+            self.check_network_convergence,
+            optimizer=optimizer,
+            fuse_all_reduce_ops=True)
+        self._compare_result_with_origin_model(
             check_func_1,
             check_func_2,
             use_cuda=False,
-            iter=2,
+            rm_drop_out=True,
+            rm_bn=True)
+        self._compare_result_with_origin_model(
+            check_func_1,
+            check_func_2,
+            use_cuda=True,
+            rm_drop_out=True,
+            rm_bn=True,
             delta2=1e-3)
 
 
diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py
index aacc1c3ecda8c25dec9f08827a856d38c37b1b2f..8960cbcdd2f574a647229894c44c2b6ea188b7d4 100644
--- a/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py
+++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py
@@ -175,7 +175,7 @@ class TestTransformer(TestParallelExecutorBase):
             self.check_network_convergence(transformer, use_cuda=True)
             self.check_network_convergence(
                 transformer, use_cuda=True, enable_sequential_execution=True)
-        self.check_network_convergence(transformer, use_cuda=False, iter=5)
+        self.check_network_convergence(transformer, use_cuda=False, iter=2)
 
 
 if __name__ == '__main__':
diff --git a/python/paddle/fluid/tests/unittests/test_recurrent_op.py b/python/paddle/fluid/tests/unittests/test_recurrent_op.py
index 6dfc85e301a2eda66bade09a8b6dd0004155f385..cf86ebf0a81c5c6cd36a5edb5d61a11cdd98ae11 100644
--- a/python/paddle/fluid/tests/unittests/test_recurrent_op.py
+++ b/python/paddle/fluid/tests/unittests/test_recurrent_op.py
@@ -15,7 +15,7 @@
 from __future__ import print_function
 
 import unittest
-
+import paddle.fluid as fluid
 import paddle.fluid.layers as layers
 from paddle.fluid.framework import Program, grad_var_name
 from paddle.fluid.executor import Executor
@@ -115,10 +115,6 @@ class RecurrentOpTest1(unittest.TestCase):
     def setup_program(self):
         self.main_program = Program()
         self.startup_program = Program()
-        self.p_info = {
-            "main_program": self.main_program,
-            "startup_program": self.startup_program
-        }
         self.place = core.CPUPlace()
 
     def setUp(self):
@@ -129,33 +125,29 @@ class RecurrentOpTest1(unittest.TestCase):
         self.output_shape = (self.sent_len, self.batch_size, self.input_dim)
         self.py_rnn = PySimpleRNN1(self.input_shape, self.output_shape)
 
-        self.output = layers.mean(self.create_rnn_op(), **self.p_info)
+        with fluid.program_guard(self.main_program, self.startup_program):
+            self.output = layers.mean(self.create_rnn_op())
 
     def create_rnn_op(self):
         x = layers.data(
             shape=[self.sent_len, self.batch_size, self.input_dim],
             dtype='float32',
             name='x',
-            append_batch_size=False,
-            **self.p_info)
+            append_batch_size=False)
         x.stop_gradient = False
         h_boot = layers.data(
-            shape=[self.input_dim],
-            dtype='float32',
-            name='h_boot',
-            **self.p_info)
+            shape=[self.input_dim], dtype='float32', name='h_boot')
         h_boot.stop_gradient = False
 
-        rnn = layers.StaticRNN(main_program=self.main_program)
+        rnn = layers.StaticRNN()
         with rnn.step():
             h_pre = rnn.memory(init=h_boot)
             x_t = rnn.step_input(x)
 
             h = layers.scale(
                 x=layers.elementwise_add(
-                    x=h_pre, y=x_t, **self.p_info),
-                scale=self.py_rnn.scale,
-                **self.p_info)
+                    x=h_pre, y=x_t),
+                scale=self.py_rnn.scale)
 
             rnn.update_memory(h_pre, h)
             rnn.output(h)
@@ -193,7 +185,8 @@ class RecurrentOpTest1(unittest.TestCase):
     def test_backward(self):
         self.check_forward()
 
-        append_backward(self.output)
+        with fluid.program_guard(self.main_program, self.startup_program):
+            append_backward(self.output)
 
         ana_grad = [np.array(x) for x in self.backward()]
 
@@ -205,12 +198,8 @@ class RecurrentOpTest1(unittest.TestCase):
                     num_grad[idx], ana_grad[idx], rtol=0.1).all())
 
     def check_forward(self):
-        print('test recurrent op forward')
         pd_output = self.forward()
         py_output = self.py_rnn.forward()
-        print('pd_output', pd_output)
-        print
-        print('py_output', py_output)
         self.assertEqual(pd_output.shape, py_output.shape)
         self.assertTrue(np.isclose(pd_output, py_output, rtol=0.1).all())
 
@@ -263,24 +252,21 @@ class RecurrentOpTest2(RecurrentOpTest1):
         self.output_shape = (self.sent_len, self.batch_size, self.input_dim)
         self.py_rnn = PySimpleRNN2(self.input_shape, self.output_shape)
 
-        self.output = layers.mean(self.create_rnn_op(), **self.p_info)
+        with fluid.program_guard(self.main_program, self.startup_program):
+            self.output = layers.mean(self.create_rnn_op())
 
     def create_rnn_op(self):
         x = layers.data(
             shape=[self.sent_len, self.batch_size, self.input_dim],
             dtype='float32',
             name='x',
-            append_batch_size=False,
-            **self.p_info)
+            append_batch_size=False)
         x.stop_gradient = False
         h_boot = layers.data(
-            shape=[self.input_dim],
-            dtype='float32',
-            name='h_boot',
-            **self.p_info)
+            shape=[self.input_dim], dtype='float32', name='h_boot')
         h_boot.stop_gradient = False
 
-        rnn = layers.StaticRNN(main_program=self.main_program)
+        rnn = layers.StaticRNN()
         with rnn.step():
             h_pre = rnn.memory(init=h_boot)
             x_t = rnn.step_input(x)
@@ -288,18 +274,13 @@ class RecurrentOpTest2(RecurrentOpTest1):
             temp_l = layers.fc(input=x_t,
                                size=self.input_dim,
                                param_attr='W',
-                               bias_attr=False,
-                               **self.p_info)
+                               bias_attr=False)
             temp_r = layers.fc(input=h_pre,
                                size=self.input_dim,
                                param_attr='U',
-                               bias_attr=False,
-                               **self.p_info)
+                               bias_attr=False)
 
-            h = layers.sigmoid(
-                x=layers.elementwise_add(
-                    x=temp_l, y=temp_r, **self.p_info),
-                **self.p_info)
+            h = layers.sigmoid(x=layers.elementwise_add(x=temp_l, y=temp_r))
 
             rnn.update_memory(h_pre, h)
             rnn.output(h)
@@ -362,40 +343,38 @@ class RecurrentOpMultipleMemoryTest(RecurrentOpTest1):
         self.py_rnn = RecurrentOpMultipleMemoryTest.PySimpleRNN3(
             self.input_shape, self.output_shape)
 
-        self.output = layers.mean(self.create_rnn_op(), **self.p_info)
+        with fluid.program_guard(self.main_program, self.startup_program):
+            self.output = layers.mean(self.create_rnn_op())
 
     def create_rnn_op(self):
         x = layers.data(
             shape=[self.sent_len, self.batch_size, self.input_dim],
             dtype='float32',
             name='x',
-            append_batch_size=False,
-            **self.p_info)
+            append_batch_size=False)
         x.stop_gradient = False
         h_boot1 = layers.data(
             shape=[self.batch_size, self.input_dim],
             dtype='float32',
             name='h_boot1',
-            append_batch_size=False,
-            **self.p_info)
+            append_batch_size=False)
         h_boot1.stop_gradient = False
         h_boot2 = layers.data(
             shape=[self.batch_size, self.input_dim],
             dtype='float32',
             name='h_boot2',
-            append_batch_size=False,
-            **self.p_info)
+            append_batch_size=False)
         h_boot2.stop_gradient = False
 
-        rnn = layers.StaticRNN(main_program=self.main_program)
+        rnn = layers.StaticRNN()
         with rnn.step():
             h_pre1 = rnn.memory(init=h_boot1)
             h_pre2 = rnn.memory(init=h_boot2)
             x_t = rnn.step_input(x)
 
-            mem1 = layers.scale(x=h_pre1, scale=1.0, **self.p_info)
-            mem2 = layers.scale(x=h_pre2, scale=1.0, **self.p_info)
-            out = layers.sums(input=[mem1, x_t, mem2], **self.p_info)
+            mem1 = layers.scale(x=h_pre1, scale=1.0)
+            mem2 = layers.scale(x=h_pre2, scale=1.0)
+            out = layers.sums(input=[mem1, x_t, mem2])
 
             rnn.update_memory(h_pre1, mem1)
             rnn.update_memory(h_pre2, mem2)
@@ -446,23 +425,23 @@ class RecurrentOpNoMemBootTest(RecurrentOpTest1):
         self.output_shape = (self.sent_len, self.batch_size, self.input_dim)
         self.py_rnn = RecurrentOpNoMemBootTest.PySimpleRNN4(self.input_shape,
                                                             self.output_shape)
-        self.output = layers.mean(self.create_rnn_op(), **self.p_info)
-        print(self.main_program)
+
+        with fluid.program_guard(self.main_program, self.startup_program):
+            self.output = layers.mean(self.create_rnn_op())
 
     def create_rnn_op(self):
         x = layers.data(
             shape=[self.sent_len, self.batch_size, self.input_dim],
             dtype='float32',
             name='x',
-            append_batch_size=False,
-            **self.p_info)
+            append_batch_size=False)
         x.stop_gradient = False
 
-        rnn = layers.StaticRNN(main_program=self.main_program)
+        rnn = layers.StaticRNN()
         with rnn.step():
             mem_pre = rnn.memory(shape=[-1, self.input_dim], batch_ref=x)
             x_t = rnn.step_input(x)
-            mem = layers.elementwise_add(x=mem_pre, y=x_t, **self.p_info)
+            mem = layers.elementwise_add(x=mem_pre, y=x_t)
             rnn.update_memory(mem_pre, mem)
             rnn.output(mem)
 
diff --git a/python/paddle/fluid/tests/unittests/test_reduce_op.py b/python/paddle/fluid/tests/unittests/test_reduce_op.py
index 8fc8125a773543eea768783155ad152c475535b5..65fc1453d8db13ad9c85746c3bf148f898e8f788 100644
--- a/python/paddle/fluid/tests/unittests/test_reduce_op.py
+++ b/python/paddle/fluid/tests/unittests/test_reduce_op.py
@@ -91,6 +91,78 @@ class TestProdOp(OpTest):
         self.check_grad(['X'], 'Out')
 
 
+class TestAllOp(OpTest):
+    def setUp(self):
+        self.op_type = "reduce_all"
+        self.inputs = {'X': np.random.randint(0, 2, (5, 6, 10)).astype("bool")}
+        self.outputs = {'Out': self.inputs['X'].all()}
+        self.attrs = {'reduce_all': True}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestAllOpWithDim(OpTest):
+    def setUp(self):
+        self.op_type = "reduce_all"
+        self.inputs = {'X': np.random.randint(0, 2, (5, 6, 10)).astype("bool")}
+        self.attrs = {'dim': [1]}
+        self.outputs = {'Out': self.inputs['X'].all(axis=1)}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestAllOpWithKeepDim(OpTest):
+    def setUp(self):
+        self.op_type = "reduce_all"
+        self.inputs = {'X': np.random.randint(0, 2, (5, 6, 10)).astype("bool")}
+        self.attrs = {'dim': [1], 'keep_dim': True}
+        self.outputs = {
+            'Out': np.expand_dims(
+                self.inputs['X'].all(axis=1), axis=1)
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestAnyOp(OpTest):
+    def setUp(self):
+        self.op_type = "reduce_any"
+        self.inputs = {'X': np.random.randint(0, 2, (5, 6, 10)).astype("bool")}
+        self.outputs = {'Out': self.inputs['X'].any()}
+        self.attrs = {'reduce_all': True}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestAnyOpWithDim(OpTest):
+    def setUp(self):
+        self.op_type = "reduce_any"
+        self.inputs = {'X': np.random.randint(0, 2, (5, 6, 10)).astype("bool")}
+        self.attrs = {'dim': [1]}
+        self.outputs = {'Out': self.inputs['X'].any(axis=1)}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestAnyOpWithKeepDim(OpTest):
+    def setUp(self):
+        self.op_type = "reduce_any"
+        self.inputs = {'X': np.random.randint(0, 2, (5, 6, 10)).astype("bool")}
+        self.attrs = {'dim': [1], 'keep_dim': True}
+        self.outputs = {
+            'Out': np.expand_dims(
+                self.inputs['X'].any(axis=1), axis=1)
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+
 class Test1DReduce(OpTest):
     def setUp(self):
         self.op_type = "reduce_sum"
diff --git a/python/paddle/fluid/trainer_desc.py b/python/paddle/fluid/trainer_desc.py
index 380c404fb2d6a36bf3732ebbff4b6cef22f71362..c742ee002aa6c470c41d46978a4e08fc774c3152 100644
--- a/python/paddle/fluid/trainer_desc.py
+++ b/python/paddle/fluid/trainer_desc.py
@@ -28,10 +28,10 @@ class TrainerDesc(object):
         import multiprocessing as mp
         # set default thread num == cpu count
         self.proto_desc.thread_num = mp.cpu_count()
-        self.fleet_desc_ = None
-        self.device_worker_ = None
-        self.program_ = None
-        self.infer_ = False
+        self._fleet_desc = None
+        self._device_worker = None
+        self._program = None
+        self._infer = False
 
     def _set_fetch_var_and_info(self, fetch_vars, fetch_info, print_period):
         for i, v in enumerate(fetch_vars):
@@ -47,19 +47,19 @@ class TrainerDesc(object):
         self.proto_desc.thread_num = thread_num
 
     def _set_device_worker(self, device_worker):
-        self.device_worker_ = device_worker
+        self._device_worker = device_worker
 
     def _set_infer(self, infer):
-        self.infer_ = infer
+        self._infer = infer
 
     def _set_fleet_desc(self, fleet_desc):
-        self.fleet_desc_ = fleet_desc
+        self._fleet_desc = fleet_desc
 
     def _gen_trainer_desc(self):
         pass
 
     def _set_program(self, program):
-        self.program_ = program
+        self._program = program
 
     def _desc(self):
         from google.protobuf import text_format
@@ -73,13 +73,13 @@ class MultiTrainer(TrainerDesc):
 
     def _set_program(self, program):
         super(MultiTrainer, self)._set_program(program)
-        self.program_ = program
+        self._program = program
 
     def _gen_trainer_desc(self):
         super(MultiTrainer, self)._gen_trainer_desc()
         self.proto_desc.class_name = "MultiTrainer"
-        self.device_worker_._set_infer(self.infer_)
-        self.device_worker_._gen_worker_desc(self.proto_desc)
+        self._device_worker._set_infer(self._infer)
+        self._device_worker._gen_worker_desc(self.proto_desc)
 
 
 class DistMultiTrainer(TrainerDesc):
@@ -89,13 +89,13 @@ class DistMultiTrainer(TrainerDesc):
 
     def _set_program(self, program):
         super(DistMultiTrainer, self)._set_program(program)
-        self.program_ = program
+        self._program = program
 
     def _gen_trainer_desc(self):
         super(DistMultiTrainer, self)._gen_trainer_desc()
         self.proto_desc.class_name = "DistMultiTrainer"
-        if self.program_ == None:
+        if self._program == None:
             raise RuntimeError("None Program")
-        self.device_worker_._set_infer(self.infer_)
-        self.device_worker_._set_program(self.program_)
-        self.device_worker_._gen_worker_desc(self.proto_desc)
+        self._device_worker._set_infer(self._infer)
+        self._device_worker._set_program(self._program)
+        self._device_worker._gen_worker_desc(self.proto_desc)
diff --git a/python/paddle/fluid/transpiler/distribute_transpiler.py b/python/paddle/fluid/transpiler/distribute_transpiler.py
index 41e5f47976c566306ad141f655a0f6516831d690..19a1f8bf74060905ecb4b81b44f7080db79c45e4 100644
--- a/python/paddle/fluid/transpiler/distribute_transpiler.py
+++ b/python/paddle/fluid/transpiler/distribute_transpiler.py
@@ -658,6 +658,7 @@ class DistributeTranspiler(object):
                 outputs={"Out": splited_var},
                 attrs={
                     "epmap": eps,
+                    "trainer_id": self.trainer_id,
                     RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
                 })
 
@@ -669,6 +670,7 @@ class DistributeTranspiler(object):
             outputs={"Out": fetch_barrier_out},
             attrs={
                 "endpoints": self.pserver_endpoints,
+                "trainer_id": self.trainer_id,
                 RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
             })
 
@@ -791,11 +793,15 @@ class DistributeTranspiler(object):
 
         global_ops = []
 
+        # sparse grad name to param name
+        sparse_grad_to_param = []
+
         def __append_optimize_op__(op, block, grad_to_block_id, merged_var,
                                    lr_ops):
             if self._is_optimizer_op(op):
                 self._append_pserver_ops(block, op, endpoint, grad_to_block_id,
-                                         self.origin_program, merged_var)
+                                         self.origin_program, merged_var,
+                                         sparse_grad_to_param)
             elif op not in lr_ops:
                 self._append_pserver_non_opt_ops(block, op)
 
@@ -911,6 +917,7 @@ class DistributeTranspiler(object):
             "Fanin": self.trainer_num,
             "sync_mode": self.sync_mode,
             "grad_to_block_id": grad_to_block_id,
+            "sparse_grad_to_param": sparse_grad_to_param,
         }
 
         if self.has_distributed_lookup_table:
@@ -1779,7 +1786,8 @@ class DistributeTranspiler(object):
         return o4
 
     def _append_pserver_ops(self, optimize_block, opt_op, endpoint,
-                            grad_to_block_id, origin_program, merged_var):
+                            grad_to_block_id, origin_program, merged_var,
+                            sparse_grad_to_param):
         program = optimize_block.program
         pserver_block = program.global_block()
         new_inputs = collections.OrderedDict()
@@ -1863,6 +1871,12 @@ class DistributeTranspiler(object):
             outputs=outputs,
             attrs=opt_op.all_attrs())
 
+        # record sparse grad to param name
+        if new_inputs["Grad"].type == core.VarDesc.VarType.SELECTED_ROWS:
+            sparse_grad_to_param.append(
+                str(new_inputs["Grad"].name) + ":" + str(new_inputs["Param"]
+                                                         .name))
+
     def _get_pserver_grad_param_var(self, var, var_dict):
         """
         Return pserver side grad/param variable, return None