diff --git a/paddle/fluid/framework/ir/multi_devices_graph_pass/fuse_all_reduce_op_pass.cc b/paddle/fluid/framework/ir/multi_devices_graph_pass/fuse_all_reduce_op_pass.cc
index dfd275d9bc5b026ba05ec57fb49b73b6de02bfd0..0525c56f3f2de4459d86f07dc4ba60e6225d46a2 100644
--- a/paddle/fluid/framework/ir/multi_devices_graph_pass/fuse_all_reduce_op_pass.cc
+++ b/paddle/fluid/framework/ir/multi_devices_graph_pass/fuse_all_reduce_op_pass.cc
@@ -276,7 +276,7 @@ class FuseAllReduceOpPass : public ir::Pass {
                                   ir::Node::Type::kOperation),
           local_scopes, places, num_of_all_reduce, multi_nccl_ctxs);
 #elif defined(PADDLE_WITH_XPU_BKCL)
-      auto *op_handle = new details::FusedAllReduceOpHandle(
+      op_handle = new details::FusedAllReduceOpHandle(
           result->CreateEmptyNode("fused_all_reduce",
                                   ir::Node::Type::kOperation),
           local_scopes, places, num_of_all_reduce, multi_bkcl_ctxs);
diff --git a/paddle/fluid/framework/ir/multi_devices_graph_pass/multi_devices_graph_pass.cc b/paddle/fluid/framework/ir/multi_devices_graph_pass/multi_devices_graph_pass.cc
index c23d357b17ef162e2cab515273fded187c12ccfb..0c03531aa889e13828048620c61836a376dbf7c4 100644
--- a/paddle/fluid/framework/ir/multi_devices_graph_pass/multi_devices_graph_pass.cc
+++ b/paddle/fluid/framework/ir/multi_devices_graph_pass/multi_devices_graph_pass.cc
@@ -522,7 +522,7 @@ void MultiDevSSAGraphBuilderBase::CreateAllReduceOp(ir::Graph *result,
               scopes, places, grad_merge_cond_name, multi_nccl_ctxs_));
 #elif defined(PADDLE_WITH_XPU_BKCL)
       result->Get<GraphOps>(kGraphOps).emplace_back(
-          new datails::GradMergeAllReduceOpHandle(
+          new details::GradMergeAllReduceOpHandle(
               result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
               scopes, places, grad_merge_cond_name, multi_bkcl_ctxs_));
 #else
diff --git a/paddle/fluid/operators/concat_op_xpu.cc b/paddle/fluid/operators/concat_op_xpu.cc
index 0558f09a174bf768c77e4a92e30d8abe4fe20021..4ebe92801e623aef0ca4e90927e2b2d0fce4d9e7 100644
--- a/paddle/fluid/operators/concat_op_xpu.cc
+++ b/paddle/fluid/operators/concat_op_xpu.cc
@@ -36,11 +36,16 @@ class ConcatXPUKernel : public framework::OpKernel<T> {
                           "XPU donot surpport AxisTensor for now"));
     axis = ComputeAxis(static_cast<int64_t>(axis),
                        static_cast<int64_t>(ins[0]->dims().size()));
-    PADDLE_ENFORCE_GE(
-        axis, 0, platform::errors::InvalidArgument("concat: axis shoud >= 0!"));
+    PADDLE_ENFORCE_GE(axis, 0, platform::errors::InvalidArgument(
+                                   "concat: axis should be larger than or "
+                                   "equal to 0, but received axis is %d.",
+                                   axis));
     PADDLE_ENFORCE_LT(axis, ins[0]->dims().size(),
                       platform::errors::InvalidArgument(
-                          "concat: axis shoud < ins[0]->dims()!"));
+                          "concat: axis should be less than ins[0]->dims()!"
+                          "But received axis is %d, while ins[0]->dims()"
+                          "size is %d.",
+                          axis, ins[0]->dims().size()));
 
     auto place = ctx.GetPlace();
     out->mutable_data<T>(place);
@@ -151,10 +156,16 @@ class ConcatGradXPUKernel : public framework::OpKernel<T> {
       }
     }
     PADDLE_ENFORCE_GE(axis, 0, platform::errors::InvalidArgument(
-                                   "concat_grad: axis shoud >= 0!"));
-    PADDLE_ENFORCE_LT(axis, out_grad->dims().size(),
-                      platform::errors::InvalidArgument(
-                          "concat_grad: axis shoud < ins[0]->dims()!"));
+                                   "concat_grad: axis should be larger than or "
+                                   "equal to 0, but received axis is %d.",
+                                   axis));
+    PADDLE_ENFORCE_LT(
+        axis, out_grad->dims().size(),
+        platform::errors::InvalidArgument(
+            "concat_grad: axis should be less than ins[0]->dims()!"
+            "But received axis is %d, while ins[0]->dims()"
+            "size is %d.",
+            axis, out_grad->dims().size()));
 
     auto input_dims = ins[0]->dims();
     std::vector<int> split_list(n);