diff --git a/norch/csrc/cuda.cu b/norch/csrc/cuda.cu
index 6908780..fdf2cb5 100644
--- a/norch/csrc/cuda.cu
+++ b/norch/csrc/cuda.cu
@@ -152,7 +152,7 @@ __global__ void sum_tensor_cuda_kernel_axis(float* data, float* result_data, int
 
         for (int i = 0; i < strides[target_axis]; ++i) {
             int index = offset + i * strides[target_axis + 1];
-            atomicAdd(&result_data[outer_index * inner_size + inner_index], data[index]);
+            //atomicAdd(&result_data[outer_index * inner_size + inner_index], data[index]);
         }
     }
 }
diff --git a/norch/csrc/tensor.cpp b/norch/csrc/tensor.cpp
index 221ec62..129a747 100644
--- a/norch/csrc/tensor.cpp
+++ b/norch/csrc/tensor.cpp
@@ -217,7 +217,7 @@ extern "C" {
                 int target_axis_stride = tensor->strides[axis];
                 int outer_size = tensor->size / target_axis_stride;
 
-                cudaMalloc((void**)&result_data, (outer_size) * sizeof(float));
+                cudaMalloc((void**)&result_data, outer_size * sizeof(float));
             }
             sum_tensor_cuda(tensor, result_data, axis);