diff --git a/doc/build/index.rst b/doc/build/index.rst
index 2b983dceb2777e6c79ee1efaa977fef6e5c33ad6..d6d0d19e110fc35faec87da90d784a6775b9c91f 100644
--- a/doc/build/index.rst
+++ b/doc/build/index.rst
@@ -9,6 +9,7 @@ Install PaddlePaddle
     :glob:
 
     install_*
+    internal/install_from_jumbo.md
 
 Build from Source
 -----------------
diff --git a/doc/cluster/index.rst b/doc/cluster/index.rst
index cf1ea97715402ec5b5b565a295ff4c1515df2570..9062f85f98d2981b5c8dcf8149e32c2ccdac77f4 100644
--- a/doc/cluster/index.rst
+++ b/doc/cluster/index.rst
@@ -5,3 +5,4 @@ Cluster Train
   :glob:
 
   opensource/cluster_train.md
+  internal/index.md
diff --git a/doc_cn/build_and_install/index.rst b/doc_cn/build_and_install/index.rst
index e9182903c5f62b3a96c196d5ba1ebba2fd14f669..e21fc98c63dcdcda8202dad349ffe24dda62492d 100644
--- a/doc_cn/build_and_install/index.rst
+++ b/doc_cn/build_and_install/index.rst
@@ -9,7 +9,11 @@ Note: The intallation packages are still in pre-release state and your experienc
 
 .. toctree::
    :maxdepth: 1
+   :glob:
    
+   源码下载(对内) <../build/internal/download_paddle_source_zh_cn.rst>
+   使用Jumbo安装(对内) <../build/internal/install_from_jumbo.rst>
+   从源码编译安装(对内)  <../build/internal/build_from_source_zh_cn.rst>
    install/docker_install.rst 
    install/ubuntu_install.rst
    cmake/index.rst
diff --git a/doc_cn/cluster/index.rst b/doc_cn/cluster/index.rst
new file mode 100644
index 0000000000000000000000000000000000000000..25313a9635bbf567a1aedfac3c379802d601d283
--- /dev/null
+++ b/doc_cn/cluster/index.rst
@@ -0,0 +1,11 @@
+集群训练
+========
+
+* `集群训练 <../../doc/cluster/index.html>`_
+
+.. toctree::
+    :maxdepth: 2
+    :glob:
+
+    集群训练(对内) <internal/index.md>
+
diff --git a/doc_cn/index.rst b/doc_cn/index.rst
index 5f06463899f6b7b8166ff2cccd87b17817c6f5d1..6cf5588b5b34f5e80ea4c70cc364d4c6c42cce3d 100644
--- a/doc_cn/index.rst
+++ b/doc_cn/index.rst
@@ -8,7 +8,7 @@ PaddlePaddle文档
 * `用户接口 <ui/index.html>`_
 * `使用示例 <demo/index.html>`_
 * `模型配置 <../doc/ui/api/trainer_config_helpers/index.html>`_
-* `集群训练 <../doc/cluster/index.html>`_
+* `集群训练 <cluster/index.html>`_
 
 开发指南
 --------
diff --git a/paddle/gserver/evaluators/CTCErrorEvaluator.cpp b/paddle/gserver/evaluators/CTCErrorEvaluator.cpp
index d0b1c0447d23d3e7072b2ee4f8e860708eb44bb2..cd4ed19c2ca45c310032c834da4cad56fb1cbdff 100644
--- a/paddle/gserver/evaluators/CTCErrorEvaluator.cpp
+++ b/paddle/gserver/evaluators/CTCErrorEvaluator.cpp
@@ -194,8 +194,8 @@ public:
   virtual real evalImp(std::vector<Argument>& arguments) {
     CHECK_EQ(arguments.size(), (size_t)2);
     Argument output, label;
-    output.resizeAndCopyFrom(arguments[0], false);
-    label.resizeAndCopyFrom(arguments[1], false);
+    output.resizeAndCopyFrom(arguments[0], false, HPPL_STREAM_DEFAULT);
+    label.resizeAndCopyFrom(arguments[1], false, HPPL_STREAM_DEFAULT);
     hl_stream_synchronize(HPPL_STREAM_DEFAULT);
     CHECK(label.sequenceStartPositions);
     CHECK(label.ids);
diff --git a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
index 7bc5fe51813c94a4347118f1366370ec8b867e02..96b0e19880b68fa2fb0ac5ad55af6d00ed401224 100644
--- a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
+++ b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
@@ -12,7 +12,6 @@ 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/utils/Stat.h"
 #include "paddle/utils/Util.h"
 #include "paddle/utils/Flags.h"
@@ -291,6 +290,8 @@ void RecurrentGradientMachine::init(
   if (subModelConfig->evaluator_names_size() > 0) {
     evaluator_.reset(frames_[0]->makeEvaluator());
   }
+
+  targetInfoInlinkId_ = subModelConfig->target_inlinkid();
 }
 
 void RecurrentGradientMachine::resizeOrCreateFrames(int numFrames) {
@@ -325,7 +326,7 @@ void RecurrentGradientMachine::resizeOrCreateFrames(int numFrames) {
 
   for (int i = frames_.size(); i < numFrames; ++i) {
     std::unique_ptr<NeuralNetwork> frame(
-          NeuralNetwork::newNeuralNetwork(subModelName_));
+        NeuralNetwork::newNeuralNetwork(subModelName_));
     frame->init(config_, subParamInitCb);
 
     for (auto& inFrameLine : inFrameLines_) {
@@ -382,6 +383,16 @@ void RecurrentGradientMachine::forward(const std::vector<Argument>& inArgs,
   size_t numSequences = input.getNumSequences();
   const int* starts = input.sequenceStartPositions->getData(false);
   bool hasSubseq = input.hasSubseq();
+
+  // In case of !hasSubseq or targetInfoInlinkId_ == -1, all inlinks share the
+  // same inframe info
+  bool shareInlinkInfo = !hasSubseq || targetInfoInlinkId_ == -1;
+
+  // Defaultly, share info with the first inlink
+  if (shareInlinkInfo) {
+    targetInfoInlinkId_ = 0;
+  }
+
   // check hasSubseq in both config and input are the same
   CHECK_EQ(hasSubseq, inFrameLines_[0].hasSubseq);
 
@@ -394,9 +405,17 @@ void RecurrentGradientMachine::forward(const std::vector<Argument>& inArgs,
     CHECK_EQ((size_t)input1.getNumSequences(), numSequences);
     // check all inputs should have same hasSubseq flag
     CHECK_EQ(input.hasSubseq(), inFrameLines_[0].hasSubseq);
-    CHECK_EQ(input1.getBatchSize(), batchSize);
-    CHECK(std::equal(starts, starts + numSequences + 1,
-                     input1.sequenceStartPositions->getData(false)));
+
+    // if shareInlinkInfo, checks:
+    // 1. all inlinks have same number of total tokens
+    // 2. all inlinks have same number of tokens for each sentence of each
+    //    sample. If hasSubseq, one sample has multiple sentence, else, one
+    //    sample is one sentence
+    if (shareInlinkInfo) {
+      CHECK_EQ(input1.getBatchSize(), batchSize);
+      CHECK(std::equal(starts, starts + numSequences + 1,
+                       input1.sequenceStartPositions->getData(false)));
+    }
   }
 
   if (hasSubseq) {
@@ -408,19 +427,44 @@ void RecurrentGradientMachine::forward(const std::vector<Argument>& inArgs,
     for (size_t i = 1; i < inFrameLines_.size(); ++i) {
       const Argument& input1 = inFrameLines_[i].inLayer->getOutput();
       CHECK_EQ((size_t)input1.getNumSubSequences(), numSubSequences);
-      CHECK(std::equal(subStarts, subStarts + numSubSequences + 1,
-                       input1.subSequenceStartPositions->getData(false)));
+      if (shareInlinkInfo) {
+        CHECK(std::equal(subStarts, subStarts + numSubSequences + 1,
+                         input1.subSequenceStartPositions->getData(false)));
+      }
     }
   }
 
   seqLengthAndStart_.clear();
-  input.getSeqLengthAndStart(&seqLengthAndStart_, &maxSequenceLength_);
+  info_.clear();
+  info_.resize(inFrameLines_.size());
+  seqLengthAndStart_.resize(inFrameLines_.size());
+
+  {
+    AsyncGpuBlock asyncGpuBlock;
+    // if shareInlinkInfo, only calculate info of the first inlink
+    // else, calculate info for each inlink
+    if (shareInlinkInfo) {
+      input.getSeqLengthAndStart(&seqLengthAndStart_[0], &maxSequenceLength_);
+      createInFrameInfo(0, input, passType);
+    } else {
+      for (size_t i = 0; i < inFrameLines_.size(); i++) {
+        const Argument& input1 = inFrameLines_[i].inLayer->getOutput();
+        input1.getSeqLengthAndStart(&seqLengthAndStart_[i],
+                                    &maxSequenceLength_);
+        createInFrameInfo(i, input1, passType);
+      }
+    }
+
+    // inFrameLine select rows in real layer one time
+    for (size_t i = 0; i < inFrameLines_.size(); i++) {
+      int curInlinkId = shareInlinkInfo ? 0 : i;
+      selectRowsOneTime(inFrameLines_[i].inLayer, info_[curInlinkId].allIds,
+                        &(inFrameLines_[i].outArg), passType);
+    }
+  }
   resizeOrCreateFrames(maxSequenceLength_);
   resizeBootFrame(numSequences);
 
-  AsyncGpuBlock asyncGpuBlock;
-  createInFrameInfo(input, passType);
-
   for (auto& memoryFrameLine : memoryFrameLines_) {
     if (memoryFrameLine.rootAgent) {
       auto scatterAgent =
@@ -443,23 +487,29 @@ void RecurrentGradientMachine::forward(const std::vector<Argument>& inArgs,
     auto gatherAgent =
         dynamic_cast<GatherAgentLayer*>(outFrameLine.agentLayer.get());
     CHECK_NOTNULL(gatherAgent);
-    gatherAgent->copyIdAndSequenceInfo(input, info_.allIds, info_.idIndex);
+    gatherAgent->copyIdAndSequenceInfo(input, info_[targetInfoInlinkId_].allIds,
+                                       info_[targetInfoInlinkId_].idIndex);
   }
 
   for (int i = 0; i < maxSequenceLength_; ++i) {
-    int idSize = info_.idIndex[i + 1] - info_.idIndex[i];
-
+    int idSize = 0;
     // connect in_links
-    for (auto& inFrameLine : inFrameLines_) {
+    for (size_t j = 0; j < inFrameLines_.size(); ++j) {
+      // idSize denotes the sum number of tokens in each length i
+      idSize = info_[j].idIndex[i + 1] - info_[j].idIndex[i];
+      InFrameLine inFrameLine = inFrameLines_[j];
       auto scatterAgent =
           dynamic_cast<ScatterAgentLayer*>(inFrameLine.agents[i].get());
       scatterAgent->setRealLayerAndOutput(inFrameLine.inLayer,
-                                          inFrameLine.outArg, info_.allIds,
-                                          info_.idIndex[i], idSize);
+                                          inFrameLine.outArg, info_[j].allIds,
+                                          info_[j].idIndex[i], idSize);
       if (hasSubseq) {
-        int size = info_.seqStartPosIndex[i + 1] - info_.seqStartPosIndex[i];
-        scatterAgent->setSequenceStartPositions(
-            info_.sequenceStartPositions, info_.seqStartPosIndex[i], size);
+        // size: the length of subsequence
+        int size =
+            info_[j].seqStartPosIndex[i + 1] - info_[j].seqStartPosIndex[i];
+        scatterAgent->setSequenceStartPositions(info_[j].sequenceStartPositions,
+                                                info_[j].seqStartPosIndex[i],
+                                                size);
       }
     }
 
@@ -469,13 +519,16 @@ void RecurrentGradientMachine::forward(const std::vector<Argument>& inArgs,
           dynamic_cast<GatherAgentLayer*>(outFrameLine.agentLayer.get());
       gatherAgent->addRealLayer(outFrameLine.frames[i]);
     }
-
     // connect memory links
+    // Adopt info_[0].idIndex because seq which has_subseq=True
+    // doesn't support Memory with !hasSubseq bootlayer;
+    // And inlinks that !hasSubSeq must have same inlink length.
+    idSize = info_[0].idIndex[i + 1] - info_[0].idIndex[i];
     for (auto& memoryFrameLine : memoryFrameLines_) {
       NeuralNetwork::connect(
           memoryFrameLine.agents[i],
           i == 0 ? memoryFrameLine.bootLayer : memoryFrameLine.frames[i - 1],
-          idSize /*height of agent*/);
+          numSeqs_[i] /*height of agent*/);
     }
   }
 
@@ -560,62 +613,76 @@ void RecurrentGradientMachine::removeBeamSearchStatisticsCallbacks() {
  * If hasSubseq, will also create scattered sequenceStartPositions infomation
  * for all realLayer of inFrameLines one time.
 */
-void RecurrentGradientMachine::createInFrameInfo(const Argument& input,
+
+void RecurrentGradientMachine::createInFrameInfo(int inlinks_id,
+                                                 const Argument& input,
                                                  PassType passType) {
   bool hasSubseq = input.hasSubseq();
+  // numSequences: # samples(sequences) in a batch
   size_t numSequences = input.getNumSequences();
   std::vector<int> allIds;
-  info_.idIndex.clear();
-  info_.idIndex.push_back(0);  // first idIndex = 0
-  if (hasSubseq) {             // for sequenceScatterAgentLayer
+
+  numSeqs_.clear();
+  Info* inlink_info = &info_[inlinks_id];
+  inlink_info->idIndex.clear();
+  inlink_info->idIndex.push_back(0);  // first idIndex = 0
+  if (hasSubseq) {                    // for sequenceScatterAgentLayer
+    // numSubSequences : all sentences within all samples(batch)
     size_t numSubSequences = input.getNumSubSequences();
     std::vector<int> sequenceStartPositions;
-    info_.seqStartPosIndex.clear();
-    info_.seqStartPosIndex.push_back(0);  // first seqStartPosIndex = 0
+    inlink_info->seqStartPosIndex.clear();
+    inlink_info->seqStartPosIndex.push_back(0);  // first seqStartPosIndex = 0
+    // maxSequenceLength_: max number of sentences(subseq) in allsamples
     for (int i = 0; i < maxSequenceLength_; ++i) {
-      sequenceStartPositions.push_back(0);  // first element = 0
-      for (size_t j = 0; j < numSubSequences; ++j) {
-        if (std::get<3>(seqLengthAndStart_[j]) == i) {
-          int subSeqStart = std::get<1>(seqLengthAndStart_[j]);
-          int subSeqLength = std::get<0>(seqLengthAndStart_[j]);
+      sequenceStartPositions.push_back(0);            // first element = 0
+      int numSeqs = 0;
+      for (size_t j = 0; j < numSubSequences; ++j) {  // for each sentence
+        // seqLengthAndStart_[inlinks_id][j]:
+        // a 4-tuple including <subseqlen, subseqstart, seqid, subseqid>
+        if (std::get<3>(seqLengthAndStart_[inlinks_id][j]) == i) {
+          ++numSeqs;
+          // subseqstart: the cpuSubSequenceStartPositions of this subseq
+          int subSeqStart = std::get<1>(seqLengthAndStart_[inlinks_id][j]);
+          int subSeqLength = std::get<0>(seqLengthAndStart_[inlinks_id][j]);
           for (int k = subSeqStart; k < subSeqStart + subSeqLength; ++k) {
             allIds.push_back(k);
           }
           sequenceStartPositions.push_back(sequenceStartPositions.back() +
-                                              subSeqLength);
+                                           subSeqLength);
         }
       }
-      info_.idIndex.push_back(allIds.size());
-      info_.seqStartPosIndex.push_back(sequenceStartPositions.size());
+      inlink_info->idIndex.push_back(allIds.size());
+      inlink_info->seqStartPosIndex.push_back(sequenceStartPositions.size());
+      numSeqs_.push_back(numSeqs);
     }
     // inFrameLine create sequenceStartPositions one time
     CHECK_EQ(sequenceStartPositions.size(),
              maxSequenceLength_ + numSubSequences);
-    CHECK_EQ(info_.seqStartPosIndex.size(),
+    CHECK_EQ(inlink_info->seqStartPosIndex.size(),
              static_cast<size_t>(maxSequenceLength_ + 1));
-    createSeqPos(sequenceStartPositions, &info_.sequenceStartPositions);
+    createSeqPos(sequenceStartPositions, &inlink_info->sequenceStartPositions);
   } else {  // for scatterAgentLayer
     for (int i = 0; i < maxSequenceLength_; ++i) {
+      int numSeqs = 0;
       for (size_t j = 0; j < numSequences; ++j) {
-        int seqLength = std::get<0>(seqLengthAndStart_[j]);
+        int seqLength = std::get<0>(seqLengthAndStart_[inlinks_id][j]);
         if (i >= seqLength) {
           break;
         }
-        int seqStart = std::get<1>(seqLengthAndStart_[j]);
+        ++numSeqs;
+        int seqStart = std::get<1>(seqLengthAndStart_[inlinks_id][j]);
         allIds.push_back(reversed_ ? (seqStart + seqLength - 1 - i)
                                    : (seqStart + i));
       }
-      info_.idIndex.push_back(allIds.size());
+      inlink_info->idIndex.push_back(allIds.size());
+      numSeqs_.push_back(numSeqs);
     }
   }
+
   // copy and check scatterId
-  copyScattedId(allIds, &info_.allIds, input.getBatchSize());
-  CHECK_EQ(info_.idIndex.size(), static_cast<size_t>(maxSequenceLength_ + 1));
-  // inFrameLine select rows in real layer one time
-  for (auto& inFrameLine : inFrameLines_) {
-    selectRowsOneTime(inFrameLine.inLayer, info_.allIds, &inFrameLine.outArg,
-                      passType);
-  }
+  copyScattedId(allIds, &inlink_info->allIds, input.getBatchSize());
+  CHECK_EQ(inlink_info->idIndex.size(),
+           static_cast<size_t>(maxSequenceLength_ + 1));
 }
 
 /* like createInFrameInfo, but for all realLayer of memoryFrameLines*/
@@ -633,19 +700,20 @@ void RecurrentGradientMachine::createMemoryFrameInfo(
     sequenceStartPositions.push_back(0);  // first element = 0
     const int* starts = input.sequenceStartPositions->getData(false);
     for (size_t i = 0; i < numSequences; ++i) {
-      int seqId = std::get<2>(seqLengthAndStart_[i]);
+      // memory info adopt info of inlinks[0]
+      int seqId = std::get<2>(seqLengthAndStart_[0][i]);
       for (int k = starts[seqId]; k < starts[seqId + 1]; ++k) {
         allIds.push_back(k);
       }
       sequenceStartPositions.push_back(sequenceStartPositions.back() +
-                                          starts[seqId + 1] - starts[seqId]);
+                                       starts[seqId + 1] - starts[seqId]);
     }
     createSeqPos(sequenceStartPositions,
                  &(*memoryFrameLine).sequenceStartPositions);
 
   } else {  // for scatterAgentLayer
     for (size_t i = 0; i < numSequences; ++i) {
-      allIds.push_back(std::get<2>(seqLengthAndStart_[i]));
+      allIds.push_back(std::get<2>(seqLengthAndStart_[0][i]));
     }
   }
   // copy and check scatterId
@@ -699,18 +767,19 @@ size_t RecurrentGradientMachine::getGenBatchSize() {
   for (auto& memoryFrameLine : memoryFrameLines_) {
     if (!memoryFrameLine.rootLayer) continue;
     Argument& bootArg = memoryFrameLine.rootLayer->getOutput();
-    size_t batchSize = memoryFrameLine.is_sequence ?
-                       bootArg.getNumSequences() : bootArg.getBatchSize();
+    size_t batchSize = memoryFrameLine.is_sequence ? bootArg.getNumSequences()
+                                                   : bootArg.getBatchSize();
     if (numSequences) {
       CHECK_EQ(numSequences, batchSize);
     } else {
       numSequences = batchSize;
     }
   }
-  CHECK(numSequences) << "Fail to get batch size in generation. "
-    "At least one of the Memory layer MUST have a layer that is NOT in "
-    "the layer group to boot it, and this boot layer is used to "
-    "decide batch_size in generation process.";
+  CHECK(numSequences)
+      << "Fail to get batch size in generation. "
+         "At least one of the Memory layer MUST have a layer that is NOT in "
+         "the layer group to boot it, and this boot layer is used to "
+         "decide batch_size in generation process.";
   return numSequences;
 }
 
@@ -732,7 +801,9 @@ void RecurrentGradientMachine::generateSequence() {
 
   // connect boot frame memory links
   std::vector<int> ids(numSequences);
-  for (size_t i = 0; i < numSequences; ++i) { ids[i] = i; }
+  for (size_t i = 0; i < numSequences; ++i) {
+    ids[i] = i;
+  }
   for (auto& memoryFrameLine : memoryFrameLines_) {
     if (memoryFrameLine.rootAgent) {
       auto scatterAgent =
@@ -756,7 +827,8 @@ void RecurrentGradientMachine::generateSequence() {
 
   // init outArg
   size_t resultNum = generator_.config.num_results_per_sample();
-  IVector::resizeOrCreate(generator_.outArg.ids,
+  IVector::resizeOrCreate(
+      generator_.outArg.ids,
       generator_.config.max_num_frames() * numSequences * resultNum, false);
   if (resultNum > 1) {
     CHECK_LE(resultNum, static_cast<size_t>(generator_.config.beam_size()));
@@ -847,7 +919,9 @@ void RecurrentGradientMachine::oneWaySearch(size_t batchSize) {
         // path.seqId = -1 indicates end of generation
         // of an input sequence
         finalPaths[seqIds_[j]].seqId = -1;
-      } else { scatterIds.push_back(j); }
+      } else {
+        scatterIds.push_back(j);
+      }
     }
   }
 
@@ -856,13 +930,12 @@ void RecurrentGradientMachine::oneWaySearch(size_t batchSize) {
   starts[0] = 0;
   generator_.ids.clear();
   for (size_t i = 0; i < batchSize; ++i) {
-    generator_.ids.insert(generator_.ids.end(),
-                          finalPaths[i].ids.begin(),
+    generator_.ids.insert(generator_.ids.end(), finalPaths[i].ids.begin(),
                           finalPaths[i].ids.end());
     starts[i + 1] = generator_.ids.size();
     batchMachineIdVec_.insert(batchMachineIdVec_.end(),
-                             finalPaths[i].machineIdVec.begin(),
-                             finalPaths[i].machineIdVec.end());
+                              finalPaths[i].machineIdVec.begin(),
+                              finalPaths[i].machineIdVec.end());
   }
 }
 
@@ -920,9 +993,9 @@ void RecurrentGradientMachine::forwardFrame(int machineCur) {
   }
 }
 
-void RecurrentGradientMachine::singlePathExpand(
-    Path& curPath, size_t curPathId, std::vector<Path>& newPaths,
-    size_t expandWidth) {
+void RecurrentGradientMachine::singlePathExpand(Path& curPath, size_t curPathId,
+                                                std::vector<Path>& newPaths,
+                                                size_t expandWidth) {
   int calc_id =
       gDiyProbStart ? gDiyProbStart(curPath.ids.size(), curPath.ids.data()) : 0;
 
@@ -946,19 +1019,20 @@ void RecurrentGradientMachine::singlePathExpand(
     if (id == -1) break;
 
     real newLogProb = generator_.config.log_prob() ? std::log(prob) : prob;
-    Path newPath(curPath, id, newLogProb,
-                 curPathId /*machineId*/, k /*topIndex*/);
+    Path newPath(curPath, id, newLogProb, curPathId /*machineId*/,
+                 k /*topIndex*/);
     if (this->beamSearchCtrlCallbacks_) {
       if (beamSearchCtrlCallbacks_->stopDetermineCandidates(
-              newPath.seqId, newPath.ids, newPath.probHistory)) return;
+              newPath.seqId, newPath.ids, newPath.probHistory))
+        return;
     }
     // outFrameLines_.size() > 1UL
     if (dataArgsSize_) {
       newPath.machineIdVec = curPath.machineIdVec;
       newPath.machineIdVec.push_back(curPathId);
     }
-    bool atEos = eosVec[index] == 1U ||
-                 newPath.ids.size() >= (size_t)maxSequenceLength_;
+    bool atEos =
+        eosVec[index] == 1U || newPath.ids.size() >= (size_t)maxSequenceLength_;
     // adjustNewPath
     newPath.adjustProb(calc_id, atEos);
     if (this->beamSearchCtrlCallbacks_) {
@@ -966,16 +1040,18 @@ void RecurrentGradientMachine::singlePathExpand(
           newPath.seqId, newPath.ids, newPath.probHistory, &newPath.logProb);
     }
     if (!newPath.isDropable()) {
-      atEos ? finalPaths_[curPath.seqId].push_back(newPath) :
-              newPaths.push_back(newPath);
+      atEos ? finalPaths_[curPath.seqId].push_back(newPath)
+            : newPaths.push_back(newPath);
     }
   }  // for expandWidth
 
-  if (gDiyProbStop) { gDiyProbStop(calc_id); }
+  if (gDiyProbStop) {
+    gDiyProbStop(calc_id);
+  }
 }
 
-void RecurrentGradientMachine::beamExpand(
-    std::vector<Path>& paths, std::vector<Path>& newPaths) {
+void RecurrentGradientMachine::beamExpand(std::vector<Path>& paths,
+                                          std::vector<Path>& newPaths) {
   size_t candidatePathCount = paths.size();
   // idVec.size() could be larger than candidatePathCount * beam,
   // so user can drop some node customly.
@@ -988,7 +1064,7 @@ void RecurrentGradientMachine::beamExpand(
   int curSeqId = 0;
   for (size_t j = 0; j <= candidatePathCount; j++) {
     // expansions of a single sequence are all processed
-    curSeqId = (j < candidatePathCount? paths[j].seqId : curSeqId + 1);
+    curSeqId = (j < candidatePathCount ? paths[j].seqId : curSeqId + 1);
     if (prevSeqId != -1 && curSeqId != prevSeqId) {
       totalExpandCount += beamShrink(newPaths, prevSeqId, totalExpandCount);
     }
@@ -1000,11 +1076,14 @@ void RecurrentGradientMachine::beamExpand(
 }
 
 // Drop extra nodes to beam size.
-size_t RecurrentGradientMachine::beamShrink(
-    std::vector<Path>& newPaths, size_t seqId, size_t totalExpandCount) {
-  size_t minNewPathSize = std::min(getBeamSize(),
-                                   newPaths.size() - totalExpandCount);
-  if (!minNewPathSize) { return 0; }
+size_t RecurrentGradientMachine::beamShrink(std::vector<Path>& newPaths,
+                                            size_t seqId,
+                                            size_t totalExpandCount) {
+  size_t minNewPathSize =
+      std::min(getBeamSize(), newPaths.size() - totalExpandCount);
+  if (!minNewPathSize) {
+    return 0;
+  }
   std::nth_element(newPaths.begin() + totalExpandCount,
                    newPaths.begin() + totalExpandCount + minNewPathSize,
                    newPaths.end(), Path::greaterPath);
@@ -1017,11 +1096,8 @@ size_t RecurrentGradientMachine::beamShrink(
 
   // Remove the already formed paths that are relatively short
   finalPaths_[seqId].erase(
-      std::remove_if(finalPaths_[seqId].begin(),
-                     finalPaths_[seqId].end(),
-                     [&](Path& p) {
-                         return p.logProb < minPathLogProb;
-                     }),
+      std::remove_if(finalPaths_[seqId].begin(), finalPaths_[seqId].end(),
+                     [&](Path& p) { return p.logProb < minPathLogProb; }),
       finalPaths_[seqId].end());
   for (auto p : finalPaths_[seqId]) {
     if (minFinalPathLogProb_[seqId] > p.logProb) {
@@ -1030,7 +1106,7 @@ size_t RecurrentGradientMachine::beamShrink(
   }
 
   if (finalPaths_[seqId].size() >= getBeamSize() &&
-          minFinalPathLogProb_[seqId] >= maxPathLogProb) {
+      minFinalPathLogProb_[seqId] >= maxPathLogProb) {
     newPaths.resize(totalExpandCount);
     return 0;
   }
@@ -1067,7 +1143,8 @@ void RecurrentGradientMachine::fillGenOutputs() {
           // in beam search, here only reserved the top 1 generated result
           // for out_links that are not the generated word indices.
           batchMachineIdVec_.insert(batchMachineIdVec_.end(),
-              path.machineIdVec.begin(), path.machineIdVec.end());
+                                    path.machineIdVec.begin(),
+                                    path.machineIdVec.end());
         }
       }
       starts[i + 1] = generator_.ids.size();
@@ -1091,21 +1168,21 @@ void RecurrentGradientMachine::copyDataOutlinkFrame(size_t machineCur) {
 
 void RecurrentGradientMachine::createDataOutlink(
     std::vector<int>& machineIdVec) {
-  size_t seqNum = getBeamSize() > 1UL ?
-                  finalPaths_.size() : finalPaths_[0].size();
+  size_t seqNum =
+      getBeamSize() > 1UL ? finalPaths_.size() : finalPaths_[0].size();
   std::vector<int> starts(seqNum + 1, 0);
   for (size_t i = 0; i < seqNum; ++i) {
-    size_t seqLen = getBeamSize() > 1UL ? finalPaths_[i][0].ids.size() :
-                                          finalPaths_[0][i].ids.size();
+    size_t seqLen = getBeamSize() > 1UL ? finalPaths_[i][0].ids.size()
+                                        : finalPaths_[0][i].ids.size();
     starts[i + 1] = starts[i] + seqLen;
   }
 
   for (size_t i = 0; i < dataArgsSize_; i++) {
-    dataArgs_[i].concat(dataArgsFrame_[i], machineIdVec,
-                        starts, useGpu_, HPPL_STREAM_1, PASS_TEST);
+    dataArgs_[i].concat(dataArgsFrame_[i], machineIdVec, starts, useGpu_,
+                        HPPL_STREAM_1, PASS_TEST);
 
-    auto dataAgent = dynamic_cast<DataLayer*>(
-        outFrameLines_[i + 1].agentLayer.get());
+    auto dataAgent =
+        dynamic_cast<DataLayer*>(outFrameLines_[i + 1].agentLayer.get());
     CHECK_NOTNULL(dataAgent);
     dataAgent->setData(dataArgs_[i]);
   }
diff --git a/paddle/gserver/gradientmachines/RecurrentGradientMachine.h b/paddle/gserver/gradientmachines/RecurrentGradientMachine.h
index cc49d13952323db6e514ea437552d076187d91e2..d9901ad81d60ae7c30ac89e3fc34ef6c179b5dd5 100644
--- a/paddle/gserver/gradientmachines/RecurrentGradientMachine.h
+++ b/paddle/gserver/gradientmachines/RecurrentGradientMachine.h
@@ -12,7 +12,6 @@ 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 "GradientMachine.h"
@@ -101,7 +100,7 @@ public:
    * Return true if this prefix or candidate is expected to be dropped.
    */
   typedef std::function<bool(int seqId, const std::vector<int>&,
-      const std::vector<real>&)> DropCallback;
+                             const std::vector<real>&)> DropCallback;
 
   /**
     * @brief NormOrDropNodeCallback
@@ -117,7 +116,7 @@ public:
     * The fourth parameter is the probability of the whole path.
     */
   typedef std::function<void(int seqId, const std::vector<int>&,
-      std::vector<real>&, real*)> NormOrDropNodeCallback;
+                             std::vector<real>&, real*)> NormOrDropNodeCallback;
 
   /**
    * @brief Register beam search control callbacks. Used for prediction.
@@ -192,7 +191,7 @@ public:
 
     int machineId;  // index of sample in frame
     int topIndex;   // index of MaxIdLayer output in one sample
-    int seqId;  // index of sequence in batch generation
+    int seqId;      // index of sequence in batch generation
     std::vector<int> machineIdVec;
 
     /**
@@ -206,7 +205,10 @@ public:
     /**
      * @brief Path default ctor, first logProb is 0.
      */
-    Path() { logProb = 0; seqId = 0; }
+    Path() {
+      logProb = 0;
+      seqId = 0;
+    }
     explicit Path(size_t seqId) : seqId(seqId) { logProb = 0; }
 
     /**
@@ -319,21 +321,37 @@ protected:
   };
   std::vector<MemoryFrameLine> memoryFrameLines_;
 
-  // All inFrameLines and outFrameLines have the same element as follows.
+  // Each inFrameLines(inlinks) has its own info(elements) below,
+  // and all outFrameLines(outlinks) share the info with one inFrameLine,
+  // which is assigned by targetInfoInlinkId_.
   struct Info {
     IVectorPtr allIds;         // scattered id of realLayer
     std::vector<int> idIndex;  // index of allIds
     ICpuGpuVectorPtr
-        sequenceStartPositions;      // scattered sequenceStartPositions
+        sequenceStartPositions;         // scattered sequenceStartPositions
     std::vector<int> seqStartPosIndex;  // index of sequenceStartPositions
   };
-  Info info_;
+  std::vector<Info> info_;
+
+  // numSeqs_[i] is the number sequences which is longer than i (for sequence
+  // data) or has more than i subsequences (for subsequence data)
+  std::vector<int> numSeqs_;
 
-  // if no subSeq, tuple of (seqLength, seqStart, seqIndex, seqIndex)
-  // else, tuple of (subSeqLength, subSeqStart, seqIndex, subSeqIndex)
-  std::vector<std::tuple<int, int, int, int>> seqLengthAndStart_;
+  // each inlinks has a "std::vector<std::tuple<int, int, int, int>>" denotes
+  // its sequence info:
+  //  if hasSubSeq, tuple of (subSeqLength, subSeqStart, seqIndex, subSeqIndex)
+  //  else, tuple of (seqLength, seqStart, seqIndex, seqIndex)
+  std::vector<std::vector<std::tuple<int, int, int, int>>> seqLengthAndStart_;
 
-  void createInFrameInfo(const Argument& input, PassType passType);
+  // the id of inlink which share info with outlinks
+  int targetInfoInlinkId_;
+
+  /* create scattered id infomation for all realLayer of inFrameLines one time.
+  *  If hasSubseq, will also create scattered sequenceStartPositions infomation
+  *  for all realLayer of inFrameLines one time.
+  */
+  void createInFrameInfo(int inlinks_id, const Argument& input,
+                         PassType passType);
 
   void createMemoryFrameInfo(MemoryFrameLine* memoryFrameLine,
                              PassType passType);
@@ -363,6 +381,9 @@ protected:
 
   NeuralNetwork* rootNetwork_;
   bool reversed_;
+
+  // if hasSubseq: max number of sentences(subseq)in batchsize samples
+  // else: max number of tokens in batchsize samples(sentences)
   int maxSequenceLength_;
   bool useGpu_;
   bool stopBeamSearch_;
@@ -415,7 +436,7 @@ private:
    * @param machineIdVec : select a row of output matrix in each frame
    * that the generation process expanded.
    */
-  void createDataOutlink(std::vector<int> & machineIdVec);
+  void createDataOutlink(std::vector<int>& machineIdVec);
 
   /*
    * @brief used in beam search, connect previous frame to form recurrent link
diff --git a/paddle/gserver/layers/CTCLayer.cpp b/paddle/gserver/layers/CTCLayer.cpp
index db1450694ecf7608fb37790e841b967288378e1f..6b9ffc5c749fb45be567881b8e625b48e28f69b4 100644
--- a/paddle/gserver/layers/CTCLayer.cpp
+++ b/paddle/gserver/layers/CTCLayer.cpp
@@ -49,8 +49,10 @@ void CTCLayer::forward(PassType passType) {
   Layer::forward(passType);
   if (useGpu_) {
     for (size_t i = 0; i < inputLayers_.size(); i++) {
-      tmpCpuInput_[i].resizeAndCopyFrom(getInput(i), false, HPPL_STREAM_1);
+      tmpCpuInput_[i].resizeAndCopyFrom(
+          getInput(i), false, HPPL_STREAM_DEFAULT);
     }
+    hl_stream_synchronize(HPPL_STREAM_DEFAULT);
     forwardImp(tmpCpuInput_[0], tmpCpuInput_[1]);
   } else {
     forwardImp(getInput(0), getInput(1));
@@ -92,9 +94,9 @@ void CTCLayer::backward(const UpdateCallback &callback) {
   if (useGpu_) {
     backwardImp(callback, tmpCpuInput_[0], tmpCpuInput_[1]);
     const_cast<Argument&>(getInput(0)).
-            resizeAndCopyFrom(tmpCpuInput_[0], true, HPPL_STREAM_1);
+            resizeAndCopyFrom(tmpCpuInput_[0], true, HPPL_STREAM_DEFAULT);
     const_cast<Argument&>(getInput(1)).
-            resizeAndCopyFrom(tmpCpuInput_[1], true, HPPL_STREAM_1);
+            resizeAndCopyFrom(tmpCpuInput_[1], true, HPPL_STREAM_DEFAULT);
   } else {
     backwardImp(callback, getInput(0), getInput(1));
   }
diff --git a/paddle/gserver/layers/ConvOperator.cpp b/paddle/gserver/layers/ConvOperator.cpp
index d08c422764e5642816a94fc55b5b67445ffb42f7..8c72c1778451dfddbaa740921cd08cf73fe56785 100644
--- a/paddle/gserver/layers/ConvOperator.cpp
+++ b/paddle/gserver/layers/ConvOperator.cpp
@@ -248,7 +248,7 @@ void ConvOperator::forward() {
   CHECK_EQ(ins_[1]->value->getHeight(), batchSize);
   checkFilterSize(ins_[1]->value);
   Matrix::resizeOrCreate(out_->value, batchSize,
-                         outputH_ * outputW_ * numFilters_);
+                         outputH_ * outputW_ * numFilters_, false, useGpu_);
   {
     AsyncGpuBlock block;
     for (size_t batchId = 0; batchId < batchSize; ++batchId) {
diff --git a/paddle/gserver/layers/CostLayer.cpp b/paddle/gserver/layers/CostLayer.cpp
index f353afabb3b7162783fef4f9093630fb826c86cb..0f99aee03200c3834c7c27343f41f77edc5a558e 100644
--- a/paddle/gserver/layers/CostLayer.cpp
+++ b/paddle/gserver/layers/CostLayer.cpp
@@ -509,8 +509,10 @@ void HuberTwoClass::forwardImp(Matrix &output, Argument &label,
                                Matrix &cost) {
   if (useGpu_) {
     for (size_t i = 0; i < inputLayers_.size(); i++) {
-      tmpCpuInput_[i].resizeAndCopyFrom(getInput(i), false, HPPL_STREAM_1);
+      tmpCpuInput_[i].resizeAndCopyFrom(
+          getInput(i), false, HPPL_STREAM_DEFAULT);
     }
+    hl_stream_synchronize(HPPL_STREAM_DEFAULT);
   }
   forwardImpIn(output, label, cost);
 }
diff --git a/paddle/gserver/layers/SamplingIdLayer.cpp b/paddle/gserver/layers/SamplingIdLayer.cpp
index 41c1461967ae1c0ff3c4b3a11e8f7405b58f6ab9..b39c9948b53118b51090059fc554e76f94316f81 100644
--- a/paddle/gserver/layers/SamplingIdLayer.cpp
+++ b/paddle/gserver/layers/SamplingIdLayer.cpp
@@ -52,8 +52,10 @@ public:
     Layer::forward(passType);
     if (useGpu_) {
       for (size_t i = 0; i < inputLayers_.size(); i++) {
-        tmpCpuInput_[i].resizeAndCopyFrom(getInput(i), false, HPPL_STREAM_1);
+        tmpCpuInput_[i].resizeAndCopyFrom(
+            getInput(i), false, HPPL_STREAM_DEFAULT);
       }
+      hl_stream_synchronize(HPPL_STREAM_DEFAULT);
       forwardImp(tmpCpuInput_[0]);
     } else {
       forwardImp(getInput(0));
diff --git a/paddle/gserver/tests/LayerGradUtil.cpp b/paddle/gserver/tests/LayerGradUtil.cpp
index f72011ae16cb3bac73e8acd5338bd7a179da329b..552a6c5b41c7f896c52b2132578b136200967573 100644
--- a/paddle/gserver/tests/LayerGradUtil.cpp
+++ b/paddle/gserver/tests/LayerGradUtil.cpp
@@ -92,7 +92,6 @@ void testState(LayerPtr testLayer, vector<DataLayerPtr>& dataLayers,
     testLayer->forward(PASS_TEST);
     Argument out;
     out.resizeAndCopyFrom(testLayer->getOutput(), /* useGpu= */ false);
-    hl_stream_synchronize(HPPL_STREAM_DEFAULT);
     if (batchOut.value) {
       size_t dim = batchOut.value->getWidth();
       ASSERT_TRUE((bool)out.value);
@@ -220,7 +219,6 @@ void testBatchState(LayerPtr testLayer, vector<DataLayerPtr>& dataLayers,
     testLayer->forward(PASS_TEST);
     Argument out;
     out.resizeAndCopyFrom(testLayer->getOutput(), /* useGpu= */ false);
-    hl_stream_synchronize(HPPL_STREAM_DEFAULT);
     if (batchOut.value) {
       size_t dim = batchOut.value->getWidth();
       ASSERT_TRUE((bool)out.value);
diff --git a/paddle/gserver/tests/Sequence/dummy.list b/paddle/gserver/tests/Sequence/dummy.list
new file mode 100644
index 0000000000000000000000000000000000000000..0e52665e11298965df5738f69c5bcefcc8bab0f9
--- /dev/null
+++ b/paddle/gserver/tests/Sequence/dummy.list
@@ -0,0 +1 @@
+dummy_file_no_use
diff --git a/paddle/gserver/tests/rnn_data_provider.py b/paddle/gserver/tests/rnn_data_provider.py
new file mode 100644
index 0000000000000000000000000000000000000000..85a83554c5c3045d144ee0250d2808237eccc9e0
--- /dev/null
+++ b/paddle/gserver/tests/rnn_data_provider.py
@@ -0,0 +1,35 @@
+# Copyright (c) 2016 Baidu, Inc. 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 paddle.trainer.PyDataProvider2 import *
+
+data = [
+    [[[1, 3, 2], [4, 5, 2]], 0],
+    [[[0, 2], [2, 5], [0, 1, 2]], 1],
+]
+
+@provider(input_types=[integer_value_sub_sequence(10),
+                       integer_value(2)])
+def process_subseq(settings, file_name):
+    for d in data:
+        yield d
+
+@provider(input_types=[integer_value_sequence(10),
+                       integer_value(2)])
+def process_seq(settings, file_name):
+    for d in data:
+        seq = []
+        for subseq in d[0]:
+            seq += subseq
+        yield seq, d[1]
diff --git a/paddle/gserver/tests/sequenceGen.py b/paddle/gserver/tests/sequenceGen.py
index e4727e472d446b48e6001968841bfc178e34ec0c..cb83d79d78cc677d5ffeb77f5693d08da2a51668 100644
--- a/paddle/gserver/tests/sequenceGen.py
+++ b/paddle/gserver/tests/sequenceGen.py
@@ -1,6 +1,3 @@
-#!/usr/bin/env python
-#coding=utf-8
-
 # Copyright (c) 2016 Baidu, Inc. All Rights Reserved
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
diff --git a/paddle/gserver/tests/sequence_nest_rnn.conf b/paddle/gserver/tests/sequence_nest_rnn.conf
new file mode 100644
index 0000000000000000000000000000000000000000..03eef7a2175880529b2743adc3ca2aaf3622c095
--- /dev/null
+++ b/paddle/gserver/tests/sequence_nest_rnn.conf
@@ -0,0 +1,75 @@
+#edit-mode: -*- python -*-
+# Copyright (c) 2016 Baidu, Inc. 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 paddle.trainer_config_helpers import *
+
+######################## data source ################################
+define_py_data_sources2(train_list='gserver/tests/Sequence/dummy.list',
+                        test_list=None,
+                        module='rnn_data_provider',
+                        obj='process_subseq')
+
+
+settings(batch_size=2, learning_rate=0.01)
+######################## network configure ################################
+dict_dim = 10
+word_dim = 8
+hidden_dim = 8
+label_dim = 3
+
+data = data_layer(name="word", size=dict_dim)
+
+emb = embedding_layer(input=data, size=word_dim)
+
+# This hierachical RNN is designed to be equivalent to the simple RNN in
+# sequence_rnn.conf
+
+def outer_step(x):
+    outer_mem = memory(name="outer_rnn_state", size=hidden_dim)
+    def inner_step(y):
+        inner_mem = memory(name="inner_rnn_state",
+                           size=hidden_dim,
+                           boot_layer=outer_mem)
+        return fc_layer(input=[y, inner_mem],
+                        size=hidden_dim,
+                        act=TanhActivation(),
+                        bias_attr=True,
+                        name="inner_rnn_state")
+
+    inner_rnn_output = recurrent_group(
+        step=inner_step,
+        input=x)
+    last = last_seq(input=inner_rnn_output, name="outer_rnn_state")
+
+    # "return last" should also work. But currently RecurrentGradientMachine
+    # does not handle it correctly. Current implementation requires that
+    # all the out links are from sequences. However, it does not report error
+    # when the out links are not sequences.
+    return inner_rnn_output
+
+out = recurrent_group(
+    step=outer_step,
+    input=SubsequenceInput(emb))
+
+value_printer_evaluator(input=out)
+
+rep = last_seq(input=out)
+prob = fc_layer(size=label_dim,
+                input=rep,
+                act=SoftmaxActivation(),
+                bias_attr=True)
+
+outputs(classification_cost(input=prob,
+                            label=data_layer(name="label", size=label_dim)))
diff --git a/paddle/gserver/tests/sequence_rnn.conf b/paddle/gserver/tests/sequence_rnn.conf
new file mode 100644
index 0000000000000000000000000000000000000000..73e7a5935f6c3cf1857d2b952471bb969d86211a
--- /dev/null
+++ b/paddle/gserver/tests/sequence_rnn.conf
@@ -0,0 +1,57 @@
+#edit-mode: -*- python -*-
+# Copyright (c) 2016 Baidu, Inc. 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 paddle.trainer_config_helpers import *
+
+######################## data source ################################
+define_py_data_sources2(train_list='gserver/tests/Sequence/dummy.list',
+                        test_list=None,
+                        module='rnn_data_provider',
+                        obj='process_seq')
+
+
+settings(batch_size=2, learning_rate=0.01)
+######################## network configure ################################
+dict_dim = 10
+word_dim = 8
+hidden_dim = 8
+label_dim = 3
+
+data = data_layer(name="word", size=dict_dim)
+
+emb = embedding_layer(input=data, size=word_dim)
+
+def step(y):
+    mem = memory(name="rnn_state", size=hidden_dim)
+    return fc_layer(input=[y, mem],
+                    size=hidden_dim,
+                    act=TanhActivation(),
+                    bias_attr=True,
+                    name="rnn_state")
+
+out = recurrent_group(
+    step=step,
+    input=emb)
+
+value_printer_evaluator(input=out)
+
+rep = last_seq(input=out)
+prob = fc_layer(size=label_dim,
+                input=rep,
+                act=SoftmaxActivation(),
+                bias_attr=True)
+
+outputs(classification_cost(input=prob,
+                            label=data_layer(name="label", size=label_dim)))
diff --git a/paddle/gserver/tests/test_RecurrentGradientMachine.cpp b/paddle/gserver/tests/test_RecurrentGradientMachine.cpp
index 35d6ee7f4a402d198dbcd1df7b272dcd65723659..f6989e9a6463a286f5df43e951b2d79e74c3d16a 100644
--- a/paddle/gserver/tests/test_RecurrentGradientMachine.cpp
+++ b/paddle/gserver/tests/test_RecurrentGradientMachine.cpp
@@ -21,6 +21,8 @@ limitations under the License. */
 #include <paddle/trainer/TrainerInternal.h>
 #include <paddle/gserver/gradientmachines/GradientMachine.h>
 
+P_DECLARE_int32(seed);
+
 using namespace paddle;  // NOLINT
 using namespace std;  // NOLINT
 class TrainerForTest : public paddle::Trainer {
@@ -68,7 +70,9 @@ void CalCost(const string& conf, const string& dir, real* cost,
   CpuVector vecMomentum(dim);
 
   // vecW needs to be assigned, otherwise the variable is an uncertain value.
-  vecW.zeroMem();
+
+  *ThreadLocalRand::getSeed() = FLAGS_seed;
+  vecW.randnorm(0, 0.1);
 
   trainer.startTrain();
   for (int i = 0; i < num_passes; ++i) {
@@ -88,15 +92,13 @@ void CalCost(const string& conf, const string& dir, real* cost,
   rmDir(dir.c_str());
 }
 
-TEST(RecurrentGradientMachine, HasSubSequence) {
+void test(const string& conf1, const string& conf2) {
   int num_passes = 5;
   real* cost1 = new real[num_passes];
-  const string conf1 = "gserver/tests/sequence_layer_group.conf";
   const string dir1 = "gserver/tests/t1";
   CalCost(conf1, dir1, cost1, num_passes);
 
   real* cost2 = new real[num_passes];
-  const string conf2 = "gserver/tests/sequence_nest_layer_group.conf";
   const string dir2 = "gserver/tests/t2";
   CalCost(conf2, dir2, cost2, num_passes);
 
@@ -109,6 +111,17 @@ TEST(RecurrentGradientMachine, HasSubSequence) {
   delete[] cost2;
 }
 
+TEST(RecurrentGradientMachine, HasSubSequence) {
+  test("gserver/tests/sequence_layer_group.conf",
+       "gserver/tests/sequence_nest_layer_group.conf");
+}
+
+TEST(RecurrentGradientMachine, rnn) {
+  test("gserver/tests/sequence_rnn.conf",
+       "gserver/tests/sequence_nest_rnn.conf");
+}
+
+
 int main(int argc, char** argv) {
   if (paddle::version::isWithPyDataProvider()) {
     if (!paddle::version::isWithGpu()) {
diff --git a/paddle/gserver/tests/test_RecurrentLayer.cpp b/paddle/gserver/tests/test_RecurrentLayer.cpp
index 2cea190b859496cd635fc5a8d1834779537d50e6..9b933b153d158bef565c0964232525ba99b8b3d4 100644
--- a/paddle/gserver/tests/test_RecurrentLayer.cpp
+++ b/paddle/gserver/tests/test_RecurrentLayer.cpp
@@ -299,7 +299,6 @@ void checkRecurrentLayer(LayerConfig layerConfig, size_t batchSize,
   Argument& cpuInput = testCpu.dataLayer_->getOutput();
   Argument& gpuInput = testGpu.dataLayer_->getOutput();
   gpuInput.resizeAndCopyFrom(cpuInput, true);
-  hl_stream_synchronize(HPPL_STREAM_DEFAULT);
 
   const VectorPtr& cpuVec = testCpu.para_->getBuf(PARAMETER_VALUE);
   const VectorPtr& gpuVec = testGpu.para_->getBuf(PARAMETER_VALUE);
diff --git a/paddle/math/Matrix.cpp b/paddle/math/Matrix.cpp
index f3a6503d4a21ff8766f3289f8eee992d4d13045d..1b7f9ac5dac16c167dcc22930c28bc3521162b9b 100644
--- a/paddle/math/Matrix.cpp
+++ b/paddle/math/Matrix.cpp
@@ -146,6 +146,7 @@ void Matrix::resizeOrCreate(MatrixPtr& matrix, size_t height, size_t width,
   if (!matrix) {
     matrix = Matrix::create(height, width, trans, useGpu);
   } else {
+    CHECK_EQ(matrix->useGpu(), useGpu);
     matrix->resize(height, width);
   }
 }
@@ -161,6 +162,7 @@ void Matrix::resizeOrCreateSparseMatrix(MatrixPtr& matrix, size_t height,
   } else {
     CHECK(dynamic_cast<CpuSparseMatrix*>(matrix.get()) ||
           dynamic_cast<GpuSparseMatrix*>(matrix.get()));
+    CHECK_EQ(matrix->useGpu(), useGpu);
     matrix->resize(height, width, nnz, valueType, format);
   }
 }
diff --git a/paddle/math/Vector.cpp b/paddle/math/Vector.cpp
index b1a459b86aa4ff70e4e07267c8a902123f9d17c0..7553ea25e09d2f52f1f8b9205f954510b77cbfa9 100644
--- a/paddle/math/Vector.cpp
+++ b/paddle/math/Vector.cpp
@@ -800,6 +800,7 @@ void CpuGpuVectorT<T>::resizeOrCreate(size_t size, bool useGpu) {
   } else if ((!useGpu) && (!cpuVectorT_)) {
     cpuVectorT_ = VectorT<T>::create(size, false);
   } else {
+    CHECK((useGpu && gpuVectorT_) || (!useGpu && cpuVectorT_));
     this->resize(size, useGpu);
   }
 }
diff --git a/paddle/parameter/Argument.cpp b/paddle/parameter/Argument.cpp
index 8610a66452358e1b2e2a846ddfcf62a0ce99e22e..a81c72aacbed3f4077afd5d604d16c922cf07013 100644
--- a/paddle/parameter/Argument.cpp
+++ b/paddle/parameter/Argument.cpp
@@ -25,6 +25,7 @@ static void resizeAndCopy(MatrixPtr& dest, const MatrixPtr& src, bool useGpu,
     if (!dest) {
       dest = src->clone(0, 0, useGpu);
     } else {
+      CHECK_EQ(dest->useGpu(), useGpu);
       dest->resize(src->getHeight(), src->getWidth());
     }
     dest->copyFrom(*src, stream);
@@ -60,12 +61,12 @@ static void resizeAndCopy(MatrixPtr& dest, const MatrixPtr& src,
                           hl_stream_t stream = HPPL_STREAM_DEFAULT) {
   if (src) {
     CHECK_LE((size_t)startRow + copySize, src->getHeight());
-
     int height = copySize;
     int width = src->getWidth();
     if (!dest) {
       dest = src->clone(height, width, useGpu);
     } else {
+      CHECK_EQ(dest->useGpu(), useGpu);
       dest->resize(height, width);
     }
     MatrixPtr submat = src->subMatrix(startRow, copySize);
@@ -182,6 +183,11 @@ static void resizeAndCopy(SVectorPtr& dest, const SVectorPtr& src,
   }
 }
 
+void Argument::resizeAndCopyFrom(const Argument& src, bool useGpu) {
+   resizeAndCopyFrom(src, useGpu, HPPL_STREAM_DEFAULT);
+   hl_stream_synchronize(HPPL_STREAM_DEFAULT);
+}
+
 void Argument::resizeAndCopyFrom(const Argument& src, bool useGpu,
                                  hl_stream_t stream) {
   dataId = src.dataId;
@@ -199,6 +205,14 @@ void Argument::resizeAndCopyFrom(const Argument& src, bool useGpu,
   resizeAndCopy(strs, src.strs, useGpu, stream);
 }
 
+int32_t Argument::resizeAndCopyFrom(const Argument& src, int32_t startSeq,
+                                    int32_t copySize, bool useGpu) {
+    int32_t size = resizeAndCopyFrom(src, startSeq, copySize, useGpu,
+                                     HPPL_STREAM_DEFAULT);
+    hl_stream_synchronize(HPPL_STREAM_DEFAULT);
+    return size;
+}
+
 int32_t Argument::resizeAndCopyFrom(const Argument& src, int32_t startSeq,
                                     int32_t copySize, bool useGpu,
                                     hl_stream_t stream) {
diff --git a/paddle/parameter/Argument.h b/paddle/parameter/Argument.h
index c444ebaf12930e938a3a4d75541d0fbf5bbb01ac..5474a05b84101ab46e0d1826626dd59399a328a9 100644
--- a/paddle/parameter/Argument.h
+++ b/paddle/parameter/Argument.h
@@ -203,13 +203,28 @@ struct Argument {
    *   startSeq: the sample id of start
    *   copySize: how many samples need to copy
    *   return value: how many samples are copied
+   * Note that when specifying the stream explicitly in this case,
+   * synchronize should also be called somewhere after this function
    */
   int32_t resizeAndCopyFrom(const Argument& src, int32_t startSeq,
-                            int32_t copySize, bool useGpu = FLAGS_use_gpu,
-                            hl_stream_t stream = HPPL_STREAM_DEFAULT);
+                            int32_t copySize, bool useGpu, hl_stream_t stream);
 
-  void resizeAndCopyFrom(const Argument& src, bool useGpu = FLAGS_use_gpu,
-                         hl_stream_t stream = HPPL_STREAM_DEFAULT);
+  /*
+   * same with the above function, except that the stream is
+   * HPPL_STREAM_DEFAULT and synchronize is automatically called
+   * inside it
+   */
+  int32_t resizeAndCopyFrom(const Argument& src, int32_t startSeq,
+                            int32_t copySize, bool useGpu = FLAGS_use_gpu);
+
+  void resizeAndCopyFrom(const Argument& src, bool useGpu, hl_stream_t stream);
+
+  /*
+   * same with the above function, except that the stream is
+   * HPPL_STREAM_DEFAULT and synchronize is automatically called
+   * inside it
+   */
+  void resizeAndCopyFrom(const Argument& src, bool useGpu = FLAGS_use_gpu);
 
   /*
     @brief Concatenate several arguments into one and put the result into it.
@@ -240,6 +255,15 @@ struct Argument {
 
   /*
    Get Sequence Length, startPositions and max Length according to input
+   1. For sequence data:
+      Each tuple is (seq_length, seq_start, seq_id, seq_id)
+      The tuples are sorted according to seq_length or subseq_length
+      *maxSequenceLength is the maximal sequence length
+
+   2. For subsequence data:
+      Each tuple is (subseq_length, subseq_start, seq_id, subseq_id)
+      The tuples are not sorted. They are in the original order.
+      *maxSequenceLenth is the maximal number of subsequences in each sequence.
    */
   void getSeqLengthAndStart(
       std::vector<std::tuple<int, int, int, int>>* seqLengthAndStart,
diff --git a/proto/ModelConfig.proto.m4 b/proto/ModelConfig.proto.m4
index d04620d363c14923455d68734b03ef9bb3f28f78..a2b243a7869eaff120b25ece35e95be4d4284d18 100644
--- a/proto/ModelConfig.proto.m4
+++ b/proto/ModelConfig.proto.m4
@@ -452,6 +452,9 @@ message SubModelConfig {
   repeated LinkConfig out_links = 10;
 
   optional GeneratorConfig generator = 11;
+
+  // the id of inlink which share info with outlinks, used in recurrent layer group
+  optional int32 target_inlinkid = 12;
 }
 
 message ModelConfig {
diff --git a/python/paddle/trainer/config_parser.py b/python/paddle/trainer/config_parser.py
index 5b60cf8410e0cd4270f7a1713afdc83027ffba10..7a6053940178db472999ed7a0d4c49cce40680e4 100644
--- a/python/paddle/trainer/config_parser.py
+++ b/python/paddle/trainer/config_parser.py
@@ -303,7 +303,8 @@ def MakeLayerNameInSubmodel(name, submodel_name = None):
 @config_func
 def RecurrentLayerGroupWithoutOutLinksBegin(name,
                                             in_links,
-                                            seq_reversed=False):
+                                            seq_reversed=False,
+                                            target_inlinkname=""):
     global g_current_submodel
     config_assert(g_config.model_config.type == "recurrent_nn",
                   "RecurrentLayerGroup should be used only in recurrent_nn")
@@ -311,14 +312,19 @@ def RecurrentLayerGroupWithoutOutLinksBegin(name,
     SubModelBegin(name)
     g_current_submodel.is_recurrent_layer_group = True
     g_current_submodel.reversed = seq_reversed
+    g_current_submodel.target_inlinkid = -1
     in_links_count = 0
-    for link in in_links:
+    for linkid, link in enumerate(in_links):
         if isinstance(link, basestring):
             name = link
             has_subseq = False
         else:
             name = link.link_name
             has_subseq = link.has_subseq
+        # assign target_inlinkid according to target_inlinkname
+        if target_inlinkname == name:
+            g_current_submodel.target_inlinkid = linkid
+
         if in_links_count == 0:
             in_links_has_subseq = has_subseq
         else:
@@ -331,6 +337,7 @@ def RecurrentLayerGroupWithoutOutLinksBegin(name,
             SequenceScatterAgentLayer(name=name, size=layer.size)
         else:
             ScatterAgentLayer(name=name, size=layer.size)
+
         pair = g_current_submodel.in_links.add()
         pair.layer_name = layer_name
         pair.link_name = MakeLayerNameInSubmodel(name)
@@ -362,10 +369,12 @@ def RecurrentLayerGroupBegin(name,
                              in_links,
                              out_links,
                              generator=None,
+                             target_inlinkname="",
                              seq_reversed=False):
     RecurrentLayerGroupWithoutOutLinksBegin(name,
                                             in_links,
-                                            seq_reversed)
+                                            seq_reversed,
+                                            target_inlinkname)
     for link in out_links:
         RecurrentLayerGroupSetOutLink(link)