diff --git a/include/pybind11/detail/common.h b/include/pybind11/detail/common.h
index cc24a52fd..8daa33903 100644
--- a/include/pybind11/detail/common.h
+++ b/include/pybind11/detail/common.h
@@ -99,6 +99,17 @@
 #  endif
 #endif
 
+// For CUDA, GCC7, GCC8:
+// PYBIND11_NOINLINE_FORCED is incompatible with `-Wattributes -Werror`.
+// When defining PYBIND11_NOINLINE_FORCED, it is best to also use `-Wno-attributes`.
+// However, the measured shared-library size saving when using noinline are only
+// 1.7% for CUDA, -0.2% for GCC7, and 0.0% for GCC8 (using -DCMAKE_BUILD_TYPE=MinSizeRel,
+// the default under pybind11/tests).
+#if !defined(PYBIND11_NOINLINE_FORCED) && \
+    (defined(__CUDACC__) || (defined(__GNUC__) && (__GNUC__ == 7 || __GNUC__ == 8)))
+#  define PYBIND11_NOINLINE_DISABLED
+#endif
+
 // The PYBIND11_NOINLINE macro is for function DEFINITIONS.
 // In contrast, FORWARD DECLARATIONS should never use this macro:
 // https://stackoverflow.com/questions/9317473/forward-declaration-of-inline-functions
diff --git a/include/pybind11/pybind11.h b/include/pybind11/pybind11.h
index 20df6a01f..7b7b3ca71 100644
--- a/include/pybind11/pybind11.h
+++ b/include/pybind11/pybind11.h
@@ -10,11 +10,6 @@
 
 #pragma once
 
-#if defined(__CUDACC__) || (defined(__GNUC__) && (__GNUC__ == 7 || __GNUC__ == 8))
-#  pragma GCC diagnostic push
-#  pragma GCC diagnostic ignored "-Wattributes"
-#endif
-
 #include "attr.h"
 #include "gil.h"
 #include "options.h"
@@ -2383,7 +2378,3 @@ PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)
 #if defined(__GNUC__) && __GNUC__ == 7
 #    pragma GCC diagnostic pop // -Wnoexcept-type
 #endif
-
-#if defined(__CUDACC__) || (defined(__GNUC__) && (__GNUC__ == 7 || __GNUC__ == 8))
-#  pragma GCC diagnostic pop
-#endif