diff --git a/norch/csrc/cuda.cu b/norch/csrc/cuda.cu
index 91b349c..ffb93af 100644
--- a/norch/csrc/cuda.cu
+++ b/norch/csrc/cuda.cu
@@ -57,7 +57,7 @@ __host__ void add_tensor_cuda(Tensor* tensor1, Tensor* tensor2, float* result_da
     cudaDeviceSynchronize();
 }
 
-__global__ void sum_tensor_cuda_kernel(float* data, float* result_data, int size) {
+_global__ void sum_tensor_cuda_kernel(float* data, float* result_data, int size) {
     extern __shared__ float sdata[];
 
     unsigned int tid = threadIdx.x;
@@ -73,7 +73,7 @@ __global__ void sum_tensor_cuda_kernel(float* data, float* result_data, int size
 
     // Reduce within block using shared memory
     for (unsigned int s = blockDim.x / 2; s > 0; s >>= 1) {
-        if (tid < s) {
+        if (tid < s && i + s < size) {
             sdata[tid] += sdata[tid + s];
         }
         __syncthreads();
diff --git a/norch/tensor.py b/norch/tensor.py
index 0a32a61..74be89f 100644
--- a/norch/tensor.py
+++ b/norch/tensor.py
@@ -20,7 +20,6 @@ class Tensor:
 
         if data != None:
             data, shape = self.flatten(data)
-            # Adjust the path to the shared library
             self.data_ctype = (ctypes.c_float * len(data))(*data)
             self.shape_ctype = (ctypes.c_int * len(shape))(*shape)
             self.ndim_ctype = ctypes.c_int(len(shape))