diff --git a/paddle/fluid/operators/reduce_ops/reduce_max_op_npu.cc b/paddle/fluid/operators/reduce_ops/reduce_max_op_npu.cc
index e3d8d15a305a98cf8b94725df2e1e5e00ada1c46..8b9925a4c20ea54ba8577d48c0f5ae7200aba2d7 100644
--- a/paddle/fluid/operators/reduce_ops/reduce_max_op_npu.cc
+++ b/paddle/fluid/operators/reduce_ops/reduce_max_op_npu.cc
@@ -141,6 +141,9 @@ class ReduceMaxGradNPUKernel : public framework::OpKernel<T> {
     Tensor tmp_out, tmp_out_grad;
     auto tmp_out_dims_vec = x_dims_vec;
     for (auto d : reduce_dims) {
+      if (d < 0) {
+        d += x_dims_vec.size();
+      }
       tmp_out_dims_vec[d] = 1;
     }