diff --git a/third_party/amd/backend/driver.c b/third_party/amd/backend/driver.c
index 5d2355059d88..dfbf3b92bf56 100644
--- a/third_party/amd/backend/driver.c
+++ b/third_party/amd/backend/driver.c
@@ -1,10 +1,10 @@
+#define PY_SSIZE_T_CLEAN
 #define __HIP_PLATFORM_AMD__
 // clang-format off
 // hip_depreated.h needs definitions from hip_runtime.h.
 #include <hip/hip_runtime.h>
 #include <hip/hip_deprecated.h>
 // clang-format on
-#define PY_SSIZE_T_CLEAN
 #include <Python.h>
 #include <dlfcn.h>
 #include <stdio.h>
diff --git a/third_party/amd/backend/driver.py b/third_party/amd/backend/driver.py
index 6cb07a0731a3..8a5964fc4dfb 100644
--- a/third_party/amd/backend/driver.py
+++ b/third_party/amd/backend/driver.py
@@ -263,6 +263,7 @@ def format_of(ty):
     params = [f"&arg{i}" for i, ty in signature.items() if ty != "constexpr"]
     params.append("&global_scratch")
     src = f"""
+#define PY_SSIZE_T_CLEAN
 #define __HIP_PLATFORM_AMD__
 #include <hip/hip_runtime.h>
 #include <Python.h>
diff --git a/third_party/nvidia/backend/driver.c b/third_party/nvidia/backend/driver.c
index ab24f7657853..e4c95fad3964 100644
--- a/third_party/nvidia/backend/driver.c
+++ b/third_party/nvidia/backend/driver.c
@@ -1,8 +1,8 @@
+#define PY_SSIZE_T_CLEAN
 #include "cuda.h"
+#include <Python.h>
 #include <dlfcn.h>
 #include <stdbool.h>
-#define PY_SSIZE_T_CLEAN
-#include <Python.h>
 
 // Raises a Python exception and returns false if code is not CUDA_SUCCESS.
 static bool gpuAssert(CUresult code, const char *file, int line) {
diff --git a/third_party/nvidia/backend/driver.py b/third_party/nvidia/backend/driver.py
index ab423518b58c..f24bf789318a 100644
--- a/third_party/nvidia/backend/driver.py
+++ b/third_party/nvidia/backend/driver.py
@@ -227,6 +227,7 @@ def format_of(ty):
     params = [f"&arg{i}" for i, ty in signature.items() if ty != "constexpr"]
     params.append("&global_scratch")
     src = f"""
+#define PY_SSIZE_T_CLEAN
 #include \"cuda.h\"
 #include <stdbool.h>
 #include <Python.h>