[NPU] Fix softmax_with_cross_entropy (#50145)

* fix cross_entropy

* update

python/paddle/nn/functional/loss.py

@@ -254,8 +254,28 @@ def fluid_softmax_with_cross_entropy(
             #        [1.15328646])
     """
     if in_dygraph_mode():
-        if core.is_compiled_with_npu():
-            softmax, backprop, loss = _legacy_C_ops.softmax_with_cross_entropy(
+        if core.is_compiled_with_custom_device("npu"):
+            if not soft_label:
+                valid_label = (
+                    paddle.cast(label != ignore_index, dtype=label.dtype)
+                    * label
+                )
+                softmax, loss = _legacy_C_ops.softmax_with_cross_entropy(
+                    logits,
+                    valid_label,
+                    'soft_label',
+                    soft_label,
+                    'ignore_index',
+                    ignore_index,
+                    'numeric_stable_mode',
+                    numeric_stable_mode,
+                    'axis',
+                    axis,
+                    'use_softmax',
+                    True,
+                )
+            else:
+                softmax, loss = _legacy_C_ops.softmax_with_cross_entropy(
                     logits,
                     label,
                     'soft_label',
@@ -266,6 +286,8 @@ def fluid_softmax_with_cross_entropy(
                     numeric_stable_mode,
                     'axis',
                     axis,
+                    'use_softmax',
+                    True,
                 )
         else:
             softmax, loss = _C_ops.cross_entropy_with_softmax(
@@ -293,7 +315,9 @@ def fluid_softmax_with_cross_entropy(
         loss = helper.create_variable_for_type_inference(dtype=logits.dtype)
 
         outputs = {'Softmax': softmax, 'Loss': loss}
-        if core.is_compiled_with_npu() or core.is_compiled_with_mlu():
+        if core.is_compiled_with_custom_device(
+            "npu"
+        ) or core.is_compiled_with_custom_device("mlu"):
             backprop = helper.create_variable_for_type_inference(
                 dtype=logits.dtype
             )
@@ -2573,7 +2597,9 @@ def cross_entropy(
             valid_label = (
                 paddle.cast(label != ignore_index, dtype=label.dtype) * label
             )
-        if core.is_compiled_with_npu() or core.is_compiled_with_mlu():
+        if core.is_compiled_with_custom_device(
+            "npu"
+        ) or core.is_compiled_with_custom_device("mlu"):
             if not soft_label:
                 _, _, out = _legacy_C_ops.softmax_with_cross_entropy(
                     input,
@@ -2744,7 +2770,9 @@ def cross_entropy(
         out = helper.create_variable_for_type_inference(dtype=input.dtype)
 
         outputs = {'Softmax': softmax, 'Loss': out}
-        if core.is_compiled_with_npu() or core.is_compiled_with_mlu():
+        if core.is_compiled_with_custom_device(
+            "npu"
+        ) or core.is_compiled_with_custom_device("mlu"):
             backprop = helper.create_variable_for_type_inference(
                 dtype=input.dtype
             )