Fedora 43 and NVCC / Cuda13.1 error "exception specification is incompatible" rsqrt / rsqrtf

Accelerated Computing CUDA CUDA Setup and Installation
1

If you are facing errors like the following when trying to complie applications with NVCC / CUDA 13.1 on Fedora 43


/usr/include/bits/mathcalls.h(206): error: exception specification is
  incompatible with that of previous function "rsqrt" (declared at line 629
  of
  /usr/local/cuda/bin/../targets/x86_64-linux/include/crt/math_functions.h)
    extern double rsqrt (double __x) noexcept (true); extern double __rsqrt (double __x) noexcept (true);
                                     ^

  /usr/include/bits/mathcalls.h(206): error: exception specification is
  incompatible with that of previous function "rsqrtf" (declared at line 653
  of
  /usr/local/cuda/bin/../targets/x86_64-linux/include/crt/math_functions.h)
    extern float rsqrtf (float __x) noexcept (true); extern float __rsqrtf (float __x) noexcept (true);
                                    ^

Its likely due to the fact that glibc’s <math.h> declares rsqrt with noexcept(true), but CUDA’s headers declare it without noexcept (or with a different exception specification).

GCC 14 + glibc are now providing rsqrt/rsqrtf in the standard math library, and there’s a conflict with CUDA’s declarations.

The simplest path is to modify the CUDA headers for math_functions.h to match the glibc math.h headers. The following worked for me

sudo patch -b /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h << 'EOF'
--- a/math_functions.h
+++ b/math_functions.h
@@ -626,7 +626,7 @@
  *
  * \note_accuracy_double
  */
-extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double                 rsqrt(double x);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double                 rsqrt(double x) noexcept (true);
 
 /**
  * \ingroup CUDA_MATH_SINGLE
@@ -650,7 +650,7 @@
  *
  * \note_accuracy_single
  */
-extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float                  rsqrtf(float x);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float                  rsqrtf(float x) noexcept (true);
 
 #if defined(__QNX__) && !defined(_LIBCPP_VERSION)
 namespace std {
EOF

Hope its helpful for others

2

thank you very much for sharing，it is useful and helpful！

3

Note that this only works with gcc, with clang the patch fails to compile.

4

Hi there. Similar issue, similar solution: opened vim and commented lines mentioned in my error (they were slightly different). Thanks!

5

btw, this incompatibility seems to be happening due to the fact that CUDA 13 does not yet officially support GCC15. So I compiled it with GCC13:

cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_HOST_COMPILER=/usr/bin/g++-13
6

Newer toolkit versions (I think since 13) now have a special macro called __NV_IEC_60559_FUNCS_EXCEPTION_SPECIFIER to deal with this. You can try if maybe it works on clang with the following patch:

--- math_functions.h.orig	2026-03-07 12:36:19.092019479 +0100
+++ math_functions.h	2026-03-07 17:15:33.322027903 +0100
@@ -604,6 +604,11 @@ extern __DEVICE_FUNCTIONS_DECL__ __devic
 #if defined(__QNX__) && !defined(_LIBCPP_VERSION)
 } /* std */
 #endif
+#if __NV_GLIBC_PROVIDES_IEC_60559_FUNCS
+#define __NV_IEC_60559_FUNCS_EXCEPTION_SPECIFIER __THROW
+#else
+#define __NV_IEC_60559_FUNCS_EXCEPTION_SPECIFIER
+#endif
 /**
  * \ingroup CUDA_MATH_DOUBLE
  * \brief Calculate the reciprocal of the square root of the input argument.
@@ -626,7 +631,7 @@ extern __DEVICE_FUNCTIONS_DECL__ __devic
  *
  * \note_accuracy_double
  */
-extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double                 rsqrt(double x);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double                 rsqrt(double x) __NV_IEC_60559_FUNCS_EXCEPTION_SPECIFIER;
 
 /**
  * \ingroup CUDA_MATH_SINGLE
@@ -650,7 +655,8 @@ extern __DEVICE_FUNCTIONS_DECL__ __devic
  *
  * \note_accuracy_single
  */
-extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float                  rsqrtf(float x);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float                  rsqrtf(float x) __NV_IEC_60559_FUNCS_EXCEPTION_SPECIFIER;
+#undef __NV_IEC_60559_FUNCS_EXCEPTION_SPECIFIER
 
 #if defined(__QNX__) && !defined(_LIBCPP_VERSION)
 namespace std {

This is the same mechanism that is now in use for sinpi / cospi and their float variants that used to cause similar problems.

7

at least on Debian, this has been just fixed in CUDA-13.2

8

I was running F42+Cuda13.1 when I noticed 13.2 was out and was supported under F43. This is good news as F42 will be out of support in a few weeks time, so I thought I should get the upgrade done. Can’t comment on your specific issue but at least the upgrade was entirely smooth for me, except for adding the 13.2 repo, so I thought it might be useful to others to reply.

I ran:

sudo dnf upgrade --refresh

sudo dnf system-upgrade download --releasever=43

sudo dnf5 offline reboot

(as I was running f42 at this point)

When it came back F43 seemed ok.

However when I went to add the cuda 13.2 repo via the Cuda installation Guide method:

sudo dnf config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/fedora43/x86_64/cuda-fedora43.repo

it failed.

it appears the dnf command line options have changed, I had to do:

sudo dnf config-manager addrepo --from-repofile=https://developer.download.nvidia.com/compute/cuda/repos/fedora43/x86_64/cuda-fedora43.repo

I ran:

sudo dnf upgrade --refresh

That upgraded the driver and cuda installation without complaint.

Then disabled the 13.1 repo in gnome software.

reboot.

As far as I can tell all seems well (obviously I have done a great amount of testing in 1 day!)

Best of luck.

9

Fixed, thanks. It will be online soon. It was still using the Fedora 40 syntax (DNF 4).

10

Glad to help.

BTW I have not tried:

dnf config-manager add-repo --from-repofile=

from the revised installation guide, but I have only seen (in dnf help) and used:

dnf config-manager addrepo --from-repofile=

add-repo might work as a synonym but thought I would mention it, just in case it doesn’t…..

11

Thanks, working on Ubuntu as well