diff --git a/paddle/phi/kernels/gpu/check_numerics_kernel.cu b/paddle/phi/kernels/gpu/check_numerics_kernel.cu
index 3238d3ece24789e849a4c8a09e71a2f57dbbb2a4..4b516b1074ba5a43196f020195f4d872a2f2e354 100644
--- a/paddle/phi/kernels/gpu/check_numerics_kernel.cu
+++ b/paddle/phi/kernels/gpu/check_numerics_kernel.cu
@@ -496,10 +496,12 @@ void CheckNumericsKernel(const Context& ctx,
                          DenseTensor* values) {
   int dev_id = tensor.place().device;
   VLOG(6) << "op_type=" << op_type << ", var_name=" << var_name
-          << ", dev_id=gpu:" << dev_id
+          << ", dev_id=gpu:" << dev_id << ", numel=" << tensor.numel()
           << ", stack_height_limit=" << stack_height_limit
           << ", output_dir=" << output_dir;
 
+  if (tensor.numel() <= 0) return;
+
   // Print to the standard output.
   char* gpu_str_ptr = GetGpuHintStringPtr<T>(ctx, op_type, var_name, dev_id);