diff --git a/extern/cuda/src/helper_cuda.cc b/extern/cuda/src/helper_cuda.cc
index bd715d5a..8c8733f4 100644
--- a/extern/cuda/src/helper_cuda.cc
+++ b/extern/cuda/src/helper_cuda.cc
@@ -15,7 +15,6 @@
 #include <cuda_runtime.h>
 #include "helper_cuda.h"
 
-
 #ifdef _CUFFT_H_
 // cuFFT API errors
 const char *_cudaGetErrorEnum(cufftResult error) {
diff --git a/src/executor.h b/src/executor.h
index 6f93cd17..3c2801ac 100644
--- a/src/executor.h
+++ b/src/executor.h
@@ -7,6 +7,10 @@
 #pragma once
 #include "common.h"
 #include "mem/allocator.h"
+#ifdef HAS_CUDA
+#include <cuda_runtime.h>
+#include "helper_cuda.h"
+#endif
 
 namespace jittor {