diff --git a/source/lib/src/gelu.cc b/source/lib/src/gelu.cc
index c554da0578..e86faa882b 100644
--- a/source/lib/src/gelu.cc
+++ b/source/lib/src/gelu.cc
@@ -37,7 +37,7 @@ void deepmd::gelu_grad_grad_cpu(
   for (int ii = 0; ii < size; ii++) {
     const FPTYPE var1 = tanh(SQRT_2_PI * (xx[ii] + 0.044715 * xx[ii] * xx[ii] *xx[ii]));
     const FPTYPE var2 = SQRT_2_PI * (1 - var1 * var1) * (0.134145 * xx[ii] * xx[ii] + 1);
-	out[ii] = dy[ii] * dy_2[ii] * (0.134145 * SQRT_2_PI * xx[ii] * xx[ii] * (1 - var1 * var1) - SQRT_2_PI * xx[ii] * var2 * (0.134145 * xx[ii] * xx[ii] + 1) * var1 + var2);
+    out[ii] = dy[ii] * dy_2[ii] * (0.134145 * SQRT_2_PI * xx[ii] * xx[ii] * (1 - var1 * var1) - SQRT_2_PI * xx[ii] * var2 * (0.134145 * xx[ii] * xx[ii] + 1) * var1 + var2);
   }
 }
 
diff --git a/source/op/gelu_multi_device.cc b/source/op/gelu_multi_device.cc
index dc86ab6c8d..af49c4ac7e 100644
--- a/source/op/gelu_multi_device.cc
+++ b/source/op/gelu_multi_device.cc
@@ -144,6 +144,10 @@ class GeluGradGradOp : public OpKernel {
         context_output_index++, 
         x_tensor.shape(),
         &output_tensor));
+    DeviceFunctor() (
+        device,
+        context->eigen_device<Device>()
+    );
     // flat the tensors
     FPTYPE * out = output_tensor->flat<FPTYPE>().data();
     const FPTYPE * x = x_tensor.flat<FPTYPE>().data();