diff --git a/paddle/fluid/operators/matmul_v2_op_npu.cc b/paddle/fluid/operators/matmul_v2_op_npu.cc
index d3022056a47ded99e63aa05c1aca8e9b31ccc3fe..353eab5bc5264146399d6d6549a32cd085902688 100644
--- a/paddle/fluid/operators/matmul_v2_op_npu.cc
+++ b/paddle/fluid/operators/matmul_v2_op_npu.cc
@@ -135,8 +135,21 @@ class MatMulV2GradNPUKernel : public framework::OpKernel<T> {
         }
         if (dy) {
           dy->mutable_data<T>(ctx.GetPlace());
-          auto runner_dy = NpuOpRunner("BatchMatMul", {*x, *dout}, {*dy},
-                                       {{"adj_x1", true}, {"adj_x2", false}});
+          framework::Tensor dout_;
+          TensorCopySync(*dout, ctx.GetPlace(), &dout_);
+          std::vector<int> vec_dim = framework::vectorize<int>(dout_.dims());
+          std::vector<int> vec_dim_v{vec_dim[0] * vec_dim[1], vec_dim[2]};
+          dout_.Resize(framework::make_ddim(vec_dim_v));
+
+          framework::Tensor x_;
+          TensorCopySync(*x, ctx.GetPlace(), &x_);
+          std::vector<int> vec_dim_x = framework::vectorize<int>(x_.dims());
+          std::vector<int> vec_dim_x_v{vec_dim_x[0] * vec_dim_x[1],
+                                       vec_dim_x[2]};
+          x_.Resize(framework::make_ddim(vec_dim_x_v));
+          auto runner_dy =
+              NpuOpRunner("MatMul", {x_, dout_}, {*dy},
+                          {{"transpose_x1", true}, {"transpose_x2", false}});
           runner_dy.Run(stream);
         }
       }
diff --git a/python/paddle/distributed/collective.py b/python/paddle/distributed/collective.py
index e28ef1e94b1a4c6b8d5c21c2dc2a712101728271..911948293c8fff056cd06242e72d5bdcc161d264 100644
--- a/python/paddle/distributed/collective.py
+++ b/python/paddle/distributed/collective.py
@@ -927,6 +927,7 @@ def _linear(x, weight, bias=None, name=None):
     else:
         helper = LayerHelper('linear', **locals())
         dtype = x.dtype
+        assert x.ndim < 4, "X latitude is not supported greater than 3 now."
 
         check_variable_and_dtype(x, 'x', ['float16', 'float32', 'float64'],
                                  'linear')