diff --git a/paddle/gserver/gradientmachines/NeuralNetwork.cpp b/paddle/gserver/gradientmachines/NeuralNetwork.cpp
index cfa80a89365af5111746eec9599d16e37532a9f7..26cff3e67710b2f38d93572c5d58849aa94a5135 100644
--- a/paddle/gserver/gradientmachines/NeuralNetwork.cpp
+++ b/paddle/gserver/gradientmachines/NeuralNetwork.cpp
@@ -202,7 +202,7 @@ void NeuralNetwork::prefetch(const std::vector<Argument>& inArgs) {
         auto mat = dynamic_cast<SparsePrefetchRowCpuMatrix*>(
             para->getMat(PARAMETER_VALUE).get());
         para->clearGradient();
-        mat->clearIndices();
+        if (mat) mat->clearIndices();
       }
     }
   }
diff --git a/paddle/parameter/Parameter.h b/paddle/parameter/Parameter.h
index e31cbc3dee6c57851c241e117dbbd9b701db9d2c..321f4275d8e68d7d3fbbc19acf0afacf689474e5 100644
--- a/paddle/parameter/Parameter.h
+++ b/paddle/parameter/Parameter.h
@@ -65,7 +65,10 @@ public:
   size_t getSize() const { return config_.size(); }
 
   bool isFullSize() const {
-    return this->getSize() == bufs_[PARAMETER_VALUE]->getSize();
+    if (bufs_[PARAMETER_VALUE]) {
+      return this->getSize() == bufs_[PARAMETER_VALUE]->getSize();
+    }
+    return false;
   }
 
   inline bool useGpu() const { return useGpu_; }
diff --git a/paddle/pserver/ParameterClient2.cpp b/paddle/pserver/ParameterClient2.cpp
index f7e391f76324a09c203dfbbb449feb050caa8fb4..54063a809a4f9e558f8d364f5c437f2b6d98925b 100644
--- a/paddle/pserver/ParameterClient2.cpp
+++ b/paddle/pserver/ParameterClient2.cpp
@@ -65,7 +65,6 @@ void ParameterClient2::initThreads() {
     LOG(INFO) << "parallel_thread_num dosent need to set";
   }
   syncThreadPool_.reset(new SyncThreadPool(threadNum_));
-
   startThreads();
 }
 
@@ -224,6 +223,14 @@ void ParameterClient2::prepareSendData(
     request.set_cost(cost);
     request.set_batch_status(batchStatus);
     CHECK_EQ(request.blocks_size(), 0);
+    VLOG(10) << "request: trainer_id: " << request.trainer_id()
+             << " update_mode" << request.update_mode()
+             << " send_back_parameter: " << request.send_back_parameter()
+             << " send_back_parameter_type: "
+             << request.send_back_parameter_type()
+             << " num_samples: " << request.num_samples()
+             << " cost: " << request.cost()
+             << " batch_status: " << request.batch_status();
   }
   for (const auto& segments : parameterSegments) {
     const auto it = parameterMap_.find(segments.id);
@@ -251,11 +258,17 @@ void ParameterClient2::prepareSendData(
       CHECK(sendMat != nullptr) << "sendMat is nullptr";
 
       syncThreadPool_->exec([&](int tid, size_t numThreads) {
+        std::lock_guard<std::mutex> guard(sparseAutoGrowthMutex_);
         const auto& localIndices = prefetchMat->getLocalIndices();
         /// num of sparse rows
         size_t nLocalBlocks = localIndices.size();
         uint64_t beginDim = 0;
         uint64_t endDim = 0;
+
+        // FIXME(typhoonzero): let it resize first
+        prefetchMat->getLocalRow(nLocalBlocks + 1);
+        sendMat->getLocalRow(nLocalBlocks + 1);
+
         for (size_t row = 0; row < nLocalBlocks; ++row) {
           int64_t blockId = localIndices[row];  // local row -> sparse row
           int serverId = std::abs((blockId + nameHash) % serviceNum_);
@@ -275,7 +288,6 @@ void ParameterClient2::prepareSendData(
           block->set_begin_pos(row * blockSize);
           /// block len
           block->set_block_size(endDim - beginDim);
-
           if (sendingPara) {
             sendJob->parallelInputIovs[serverId].push_back(
                 {sendMat->getLocalRow(row), sizeof(real) * (size_t)blockSize});
diff --git a/paddle/pserver/ParameterClient2.h b/paddle/pserver/ParameterClient2.h
index 89b3ddd502151e537b81bdbb09f171dd6e13ba26..29b9eeacddf2945dd22b7b17fc87c7c74b868896 100644
--- a/paddle/pserver/ParameterClient2.h
+++ b/paddle/pserver/ParameterClient2.h
@@ -583,6 +583,7 @@ protected:
 #ifndef PADDLE_DISABLE_TIMER
   uint64_t forwardbackwordTime_;
 #endif
+  std::mutex sparseAutoGrowthMutex_;
 
   /// map id to parameter used for decoding protobuf data
   std::unordered_map<size_t, ParameterPtr> parameterMap_;