fix dgcclipnorm bug test=develop (#16629)

Author: gongweibao

paddle/fluid/operators/dgc_clip_by_norm_op.h

@@ -24,18 +24,21 @@ class DGCClipByNormKernel : public ClipByNormKernel<DeviceContext, T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     auto rampup_begin_step = context.Attr<float>("rampup_begin_step");
-    if (static_cast<int>(rampup_begin_step) >= 0) {
-      auto current_step_tensor =
-          context.Input<framework::Tensor>("current_step");
-      auto* current_step = current_step_tensor->data<T>();
-
-      if (static_cast<int>(*current_step) <
-          static_cast<int>(rampup_begin_step)) {
-        VLOG(10) << "current_step:" << *current_step
-                 << " < rampup_begin_step:" << rampup_begin_step
-                 << " so does't use dgc_clip_by_norm";
-        return;
-      }
+    if (static_cast<int>(rampup_begin_step) < 0) {
+      return;
+    }
+
+    auto current_step_tensor = context.Input<framework::Tensor>("current_step");
+    auto* current_step = current_step_tensor->data<T>();
+
+    VLOG(10) << "current_step:" << *current_step
+             << ", rampup_begin_step:" << rampup_begin_step;
+
+    if (static_cast<int>(*current_step) < static_cast<int>(rampup_begin_step)) {
+      VLOG(10) << "current_step:" << *current_step
+               << " < rampup_begin_step:" << rampup_begin_step
+               << " so does't use dgc_clip_by_norm";
+      return;
     }
 
     return ClipByNormKernel<DeviceContext, T>::Compute(context);

python/paddle/fluid/optimizer.py

@@ -832,7 +832,7 @@ def _clip_by_norm(self, x, max_norm, name=None):
             type=x.type, name=name, dtype=x.dtype, persistable=False)
 
         helper.append_op(
-            type="clip_by_norm",
+            type="dgc_clip_by_norm",
             inputs={"X": x,
                     "current_step": self._global_step_var},
             attrs={
@@ -845,7 +845,7 @@ def _clip_by_norm(self, x, max_norm, name=None):
     def _append_clip_norm(self, grad_var, clip_norm):
         with grad_var.block.program._backward_role_guard():
             return self._clip_by_norm(
-                x=grad_var, max_norm=clip_norm, name=grad_var.name + "@DGC")
+                x=grad_var, max_norm=clip_norm, name=grad_var.name)
 
     def _dgc_op(self, param_var, clip_var, grad_var, u_var, v_var, k_var,
                 encoded_var):