diff --git a/paddle/fluid/operators/fused/fused_multi_transformer_op.cu b/paddle/fluid/operators/fused/fused_multi_transformer_op.cu
index f4a5319a68caac23882b1b0902a07ffa20000248..e38ac9a0ad2da52c62b4f64f4ea50eaaa90faec9 100644
--- a/paddle/fluid/operators/fused/fused_multi_transformer_op.cu
+++ b/paddle/fluid/operators/fused/fused_multi_transformer_op.cu
@@ -534,6 +534,8 @@ template <typename T, int Dh, int THREADS_PER_KEY, int THREADS_PER_VALUE,
           int THREADS_PER_BLOCK>
 __global__ void masked_multihead_attention_kernel(
     Masked_multihead_attention_params<T> params) {
+#if CUDA_ARCH_FP16_SUPPORTED(__CUDA_ARCH__)
+
   static_assert(Dh % THREADS_PER_KEY == 0, "");
   static_assert(Dh % THREADS_PER_VALUE == 0, "");
 
@@ -821,6 +823,9 @@ __global__ void masked_multihead_attention_kernel(
     printf("\n");
   }
 #endif
+#else
+  assert(false);
+#endif
 }
 
 template <typename T>