diff --git a/paddle/fluid/operators/collective/c_softmax_with_cross_entropy_op.cu b/paddle/fluid/operators/collective/c_softmax_with_cross_entropy_op.cu
index 114c52f608f341..0e1d4640b93aeb 100644
--- a/paddle/fluid/operators/collective/c_softmax_with_cross_entropy_op.cu
+++ b/paddle/fluid/operators/collective/c_softmax_with_cross_entropy_op.cu
@@ -188,8 +188,7 @@ struct CSoftmaxWithCrossEntropyFunctor<phi::GPUContext, T> {
 
     eigen_softmax.device(*dev_ctx.eigen_device()) =
         (eigen_logits -
-         eigen_logits_max.reshape(batch_by_one).broadcast(one_by_class))
-            .unaryExpr(phi::funcs::ValueClip<T>());
+         eigen_logits_max.reshape(batch_by_one).broadcast(one_by_class));
 
     // step 3, obtain predict target
     phi::DenseTensor predicted_logits;
@@ -346,8 +345,7 @@ struct CSoftmaxWithCrossEntropyProcessGroupFunctor<phi::GPUContext, T> {
 
     eigen_softmax.device(*dev_ctx.eigen_device()) =
         (eigen_logits -
-         eigen_logits_max.reshape(batch_by_one).broadcast(one_by_class))
-            .unaryExpr(phi::funcs::ValueClip<T>());
+         eigen_logits_max.reshape(batch_by_one).broadcast(one_by_class));
 
     // step 3, obtain predict target
     phi::DenseTensor predicted_logits;
diff --git a/python/paddle/fluid/tests/unittests/c_softmax_with_cross_entropy_op.py b/python/paddle/fluid/tests/unittests/c_softmax_with_cross_entropy_op.py
index 0217f7ee710f53..684c1c20bb1d7d 100644
--- a/python/paddle/fluid/tests/unittests/c_softmax_with_cross_entropy_op.py
+++ b/python/paddle/fluid/tests/unittests/c_softmax_with_cross_entropy_op.py
@@ -26,7 +26,7 @@ def stable_softmax(x):
     """Compute the softmax of vector x in a numerically stable way."""
     # clip to shiftx, otherwise, when calc loss with
     # log(exp(shiftx)), may get log(0)=INF
-    shiftx = (x - np.max(x)).clip(-64.0)
+    shiftx = x - np.max(x)
     exps = np.exp(shiftx)
     return exps / np.sum(exps)
 
@@ -88,13 +88,13 @@ def test_model(self, data_type="float32"):
         # get input data for rank 0
         np.random.seed(0)
         input0 = np.random.uniform(
-            low=-10.0, high=10.0, size=(self.batch_size, local_elements)
+            low=-40.0, high=40.0, size=(self.batch_size, local_elements)
         ).astype(data_type)
 
         # get input data for rank 1
         np.random.seed(1)
         input1 = np.random.uniform(
-            low=-10.0, high=10.0, size=(self.batch_size, local_elements)
+            low=-40.0, high=40.0, size=(self.batch_size, local_elements)
         ).astype(data_type)
 
         # get combined input data