diff --git a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
index f98bf95064fa539b990309dfe0bff10c1e99d096..157b1ab45163a94a81d859dbcb7a52ae8edae439 100644
--- a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
+++ b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
@@ -184,7 +184,7 @@ public:
   }
 
   void backward(const UpdateCallback& callback) override {
-    if (biases_) {
+    if (biases_ && biases_->getWGrad()) {
       backwardActivation();
       biases_->getWGrad()->collectBias(*getOutputGrad(), 1);
       biases_->getParameterPtr()->incUpdate(callback);