diff --git a/paddle/fluid/operators/collective/c_softmax_with_cross_entropy_op_xpu.cc b/paddle/fluid/operators/collective/c_softmax_with_cross_entropy_op_xpu.cc
index a1b9f3d9d8bd428d4a386fe7c98de7367fbdcc0d..83824c810840189502105ca3dc95dbd42713d4c3 100644
--- a/paddle/fluid/operators/collective/c_softmax_with_cross_entropy_op_xpu.cc
+++ b/paddle/fluid/operators/collective/c_softmax_with_cross_entropy_op_xpu.cc
@@ -144,6 +144,12 @@ struct CSoftmaxWithCrossEntropyProcessGroupFunctor<phi::XPUContext, T> {
     phi::DenseTensor predicted_logits;
     predicted_logits =
         ctx.AllocateTmpTensor<T, phi::XPUContext>({N, 1}, dev_ctx);
+    ret = xpu::constant<XPUType>(
+        dev_ctx.x_context(),
+        reinterpret_cast<XPUType*>(predicted_logits.data<T>()),
+        N,
+        0.0);
+    PADDLE_ENFORCE_XDNN_SUCCESS(ret, "constant");
     const int start_index = rank * D;
     const int end_index = start_index + D;
     const auto& label_type = framework::TransToProtoVarType(labels->dtype());