diff --git a/onnxruntime/core/providers/cuda/tensor/slice_impl.cu b/onnxruntime/core/providers/cuda/tensor/slice_impl.cu
index f98bc19837..0c5b3eb514 100644
--- a/onnxruntime/core/providers/cuda/tensor/slice_impl.cu
+++ b/onnxruntime/core/providers/cuda/tensor/slice_impl.cu
@@ -29,12 +29,15 @@ __global__ void _SliceKernel(const int32_t dimension_count,
       break;
     }
 
-    output_strides[dim].divmod(value, div, mod);
-    input_index += (starts[dim] + div * steps[dim]) * input_strides[dim];
+    output_strides.data_[dim].divmod(value, div, mod);
+    input_index += (starts.data_[dim] + div * steps.data_[dim]) * input_strides.data_[dim];
     value = mod;
   }
-  input_index += starts[dim] + mod * steps[dim];
-  output_data[id] = input_data[input_index];
+  input_index += starts.data_[dim] + mod * steps.data_[dim];
+  if (is_grad)
+    output_data[input_index] = input_data[id];
+  else
+    output_data[id] = input_data[input_index];
 }
 
 Status SliceImpl(const size_t element_size,