diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc
index b5dad398448f75..fd8efe3c6f0a85 100644
--- a/paddle/fluid/framework/operator.cc
+++ b/paddle/fluid/framework/operator.cc
@@ -3230,6 +3230,29 @@ void OperatorWithKernel::BuildPhiKernelContext(
   }
   VLOG(4) << "Done attributes";
 
+// Clear All old attrs before add new attrs,
+// because sometimes old attrs may be misused.
+#if defined(PADDLE_WITH_MKLDNN)
+  if (phi::OneDNNContext::classof(dev_ctx)) {
+    phi::OneDNNContext* one_dnn_ctx = static_cast<phi::OneDNNContext*>(dev_ctx);
+    one_dnn_ctx->ClearDnnAttr();
+  }
+#endif
+
+  // Note(YuanRisheng): Now, we can't open code below.
+  // Because some unittest run OLD dygraph and ExtraAttr is not supported in OLD
+  // dygraph. So, here we use trick that dev_ctx is a global object. We can
+  // store ExtraAttr in static graph and when unittest run OLD dygraph, it can
+  // obtain these ExtraAttr. We can open this code when OLD dygraph is no longer
+  // used.
+  /*
+  #if defined(PADDLE_WITH_CUDA)
+    if(phi::GPUContext::classof(dev_ctx)) {
+      phi::GPUContext* gpu_dnn_ctx = static_cast<phi::GPUContext*>(dev_ctx);
+      gpu_dnn_ctx->ClearDnnAttr();
+    }
+  #endif
+  */
   // For compatible with Op with extra attrs for specific backend
 #if defined(PADDLE_WITH_MKLDNN) || defined(PADDLE_WITH_CUDA)
   auto& runtime_attrs = RuntimeAttrs();
diff --git a/paddle/phi/backends/gpu/gpu_context.cc b/paddle/phi/backends/gpu/gpu_context.cc
index 0e102911442f2e..e4dcf3908f957a 100644
--- a/paddle/phi/backends/gpu/gpu_context.cc
+++ b/paddle/phi/backends/gpu/gpu_context.cc
@@ -740,6 +740,8 @@ struct GPUContext::Impl {
     dnn_attrs_[attr_name] = attr;
   }
 
+  void ClearDnnAttr() { dnn_attrs_.clear(); }
+
   // use one flag for all handles?
   // they should be accessed consistently
   bool owned_{false};
@@ -1042,4 +1044,6 @@ void GPUContext::SetDnnAttr(const std::string& attr_name, Attribute attr) {
   return impl_->SetDnnAttr(attr_name, std::move(attr));
 }
 
+void GPUContext::ClearDnnAttr() { return impl_->ClearDnnAttr(); }
+
 }  // namespace phi
diff --git a/paddle/phi/backends/gpu/gpu_context.h b/paddle/phi/backends/gpu/gpu_context.h
index 84aba73fad1e12..0b34d95eaf0d21 100644
--- a/paddle/phi/backends/gpu/gpu_context.h
+++ b/paddle/phi/backends/gpu/gpu_context.h
@@ -172,6 +172,7 @@ class PADDLE_API GPUContext : public DeviceContext,
   bool HasDnnAttr(const std::string& attr_name) const;
   const Attribute& GetDnnAttr(const std::string& attr_name) const;
   void SetDnnAttr(const std::string& attr_name, Attribute attr);
+  void ClearDnnAttr();
 
   static const char* name() { return "GPUContext"; }
 
diff --git a/paddle/phi/backends/onednn/onednn_context.cc b/paddle/phi/backends/onednn/onednn_context.cc
index 6c69191c944573..2678506f0b9a84 100644
--- a/paddle/phi/backends/onednn/onednn_context.cc
+++ b/paddle/phi/backends/onednn/onednn_context.cc
@@ -301,6 +301,8 @@ struct OneDNNContext::Impl {
     dnn_attrs_[attr_name] = attr;
   }
 
+  void ClearDnnAttr() { dnn_attrs_.clear(); }
+
   bool HasDnnInput(const std::string& input_name) const {
     return dnn_inputs_.count(input_name) != 0UL;
   }
@@ -425,6 +427,8 @@ void OneDNNContext::SetDnnAttr(const std::string& attr_name, Attribute attr) {
   return impl_->SetDnnAttr(attr_name, std::move(attr));
 }
 
+void OneDNNContext::ClearDnnAttr() { return impl_->ClearDnnAttr(); }
+
 bool OneDNNContext::HasDnnInput(const std::string& input_name) const {
   return impl_->HasDnnInput(input_name);
 }
diff --git a/paddle/phi/backends/onednn/onednn_context.h b/paddle/phi/backends/onednn/onednn_context.h
index 9035aef5f9a328..79eaa05948c622 100644
--- a/paddle/phi/backends/onednn/onednn_context.h
+++ b/paddle/phi/backends/onednn/onednn_context.h
@@ -146,6 +146,8 @@ class OneDNNContext : public CPUContext {
   const DenseTensor* GetDnnInput(const std::string& input_name) const;
   void SetDnnInput(const std::string& input_name, const DenseTensor* input);
 
+  void ClearDnnAttr();
+
   void SetInputsName(const TensorNameMap& inputs_name);
 
   void SetOutputsName(const TensorNameMap& outputs_name);