diff --git a/paddle/operators/clip_op.cu b/paddle/operators/clip_op.cu
index 883f25c9271558556430c3424df7feb8a25c24e1..ac6a062f6d9fc84b43d1fe74c0d4e922f16be494 100644
--- a/paddle/operators/clip_op.cu
+++ b/paddle/operators/clip_op.cu
@@ -54,12 +54,11 @@ class ClipGradientOpCUDAKernel : public framework::OpKernel {
     int D = d_x->dims()[1];
     int block = 512;
     int grid = (N * D + block - 1) / block;
-    auto* device_context =
-        const_cast<platform::DeviceContext*>(context.device_context_);
-    ClipGradientKernel<
-        T><<<grid, block, 0,
-             reinterpret_cast<platform::CUDADeviceContext*>(device_context)
-                 ->stream()>>>(count, min, max, x_data, d_out_data, d_x_data);
+    ClipGradientKernel<T><<<
+        grid, block, 0, reinterpret_cast<const platform::CUDADeviceContext&>(
+                            context.device_context())
+                            .stream()>>>(count, min, max, x_data, d_out_data,
+                                         d_x_data);
   }
 };