[Paddle Inference] Set softmax op use_cudnn default true. (#56036)

* fix_softmax_eigen

* fix_ctest_seresnet

* fix_ci_error

paddle/phi/api/yaml/op_compat.yaml

@@ -2528,7 +2528,7 @@
     softmax : GetSoftmaxExpectedKernelType
     softmax_grad : GetSoftmaxGradExpectedKernelType
   extra :
-    attrs : [str data_format = "AnyLayout", bool use_cudnn=false, bool use_mkldnn = false, str mkldnn_data_type = "float32", bool is_test = false]
+    attrs : [str data_format = "AnyLayout", bool use_cudnn = true, bool use_mkldnn = false, str mkldnn_data_type = "float32", bool is_test = false]
 
 - op : softplus
   backward : softplus_grad, softplus_double_grad

test/legacy_test/dist_se_resnext.py

@@ -218,7 +218,7 @@ class DistSeResneXt2x2(TestDistRunnerBase):
         model = SE_ResNeXt(layers=50)
         out = model.net(input=image, class_dim=102)
         cost = paddle.nn.functional.cross_entropy(
-            input=out, label=label, reduction='none', use_softmax=False
+            input=out, label=label, reduction='none', use_softmax=True
         )
 
         avg_cost = paddle.mean(x=cost)

test/legacy_test/seresnext_net.py

@@ -171,7 +171,7 @@ def SE_ResNeXt50Small(use_feed):
     # Classifier layer:
     prediction = paddle.static.nn.fc(x=dropout, size=1000, activation='softmax')
     loss = paddle.nn.functional.cross_entropy(
-        input=prediction, label=label, reduction='none', use_softmax=False
+        input=prediction, label=label, reduction='none', use_softmax=True
     )
     loss = paddle.mean(loss)
     return loss