Fix compilation on Pop!_OS 22.04 LTS CUDA

[mika314]

Downgrade the C++ version from 17 to 14, and also update the code to remove compilation warnings.

Without that fix the build fails with the following error.

/usr/include/c++/11/bits/std_function.h:435:145: error: parameter packs not expanded with ‘...’:
  435 |         function(_Functor&& __f)
      |                                                                                                                                                 ^ 
/usr/include/c++/11/bits/std_function.h:435:145: note:         ‘_ArgTypes’
/usr/include/c++/11/bits/std_function.h:530:146: error: parameter packs not expanded with ‘...’:
  530 |         operator=(_Functor&& __f)
      |                                                                                                                                                  ^ 
/usr/include/c++/11/bits/std_function.h:530:146: note:         ‘_ArgTypes’
gmake[2]: *** [ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make:107: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.o] Error 1

Make sure to read the contributing guidelines before submitting a PR

[max-krasnyansky]

Making C++ standard conditional on some backend is not a good idea. We recently decided to go C++17 by default.

PopOS is based on Ubuntu. Why don't you just get new llvm-18 or -19 stack installed and build with llvm/clang.

wget https://apt.llvm.org/llvm.sh
chmod +x llvm.sh
sudo ./llvm.sh 18

Then build llama.cpp with

CC=clang-18 CXX=clang++-18 cmake ...

[mika314]

I did not know I can build it with clang. Clang should work with C++17 no problem. I'll try to rebuild it with clang.

Before I try, one question. I suspect the problem might be with CUDA because it is failing to build .cu file?

[mika314]

I played around with nvcc a bit and reprod the issue.

#include <iostream>
#include <functional>

__global__ void hello() {
    if constexpr (true) {
        printf("C++17 is supported!\n");
    }
}

int main() {
    hello<<<1, 1>>>();
    cudaDeviceSynchronize();
    return 0;
}

$ nvcc -std=c++17 -o main main.cu /usr/include/c++/11/bits/std_function.h:435:145: error: parameter packs not expanded with ‘...’: 435 | function(_Functor&& __f) | ^ /usr/include/c++/11/bits/std_function.h:435:145: note: ‘_ArgTypes’ /usr/include/c++/11/bits/std_function.h:530:146: error: parameter packs not expanded with ‘...’: 530 | operator=(_Functor&& __f) | ^ /usr/include/c++/11/bits/std_function.h:530:146: note: ‘_ArgTypes’

I suspect using clang would not help. I need to upgrade nvcc.

$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2021 NVIDIA Corporation
Built on Thu_Nov_18_09:45:30_PST_2021
Cuda compilation tools, release 11.5, V11.5.119
Build cuda_11.5.r11.5/compiler.30672275_0

Maybe there is a way build nvcc on C++14 and rest on C++17. I am not expert in cmake.

[max-krasnyansky]

Ah. Yeah, sorry, my bad. You need newer CUDA toolkit. CUDA 12 should work (it does on Ubuntu 24.04). https://toranbillups.com/blog/archive/2023/08/19/install-cuda-12-on-popos/

Technically you can explicitly specify the host compiler to use (NVCC_PREPEND_FLAGS) but it seems to be picky about the specific compiler versions it supports.

[mika314]

Thanks for the suggestion. I'll try to update to CUDA 12.