Fix CUDA toolchain detection for Arch builds

[IldarMinaev]

Summary

Fix CUDA-enabled Arch builds by making the CUDA toolkit/compiler discovery explicit and by
propagating CUDA toolchain settings into the nested whisper.cpp CMake build.

This changes:

• arch/git/PKGBUILD and arch/release/PKGBUILD
  • discover nvcc from CUDACXX, CUDAToolkit_ROOT, CUDA_PATH, CUDA_HOME, PATH, and finally the Arch
    package fallback /opt/cuda/bin/nvcc;
  • export CUDA_PATH, CUDAToolkit_ROOT, and the directory containing nvcc before enabling
    BUILD_WHISPERCPP_CUBLAS;
  • pass CUDAToolkit_ROOT, CMAKE_CUDA_COMPILER, and, when available,
    CMAKE_CUDA_HOST_COMPILER;
  • honor user-provided CUDAHOSTCXX / NVCC_CCBIN before falling back to Arch's g++-15.
• cmake/whispercpp.cmake
  • forwards CUDA toolkit/compiler settings to the whispercppcublas ExternalProject;
  • sets CMAKE_CUDA_STANDARD=17 for the nested CUDA build.

Problem

AUR package dsnote-git currently enables BUILD_WHISPERCPP_CUBLAS on Arch when the cuda package is installed:

pacman -Qi cuda &> /dev/null && HAS_CUDA=true || HAS_CUDA=false

That check is not sufficient for a non-interactive package build. Arch installs CUDA under /opt/cuda, and the
CUDA package's profile script adds /opt/cuda/bin to PATH and sets NVCC_CCBIN=/usr/bin/g++-15. Package build
environments are not guaranteed to have that profile state loaded.

The result is that BUILD_WHISPERCPP_CUBLAS=ON can be enabled while the nested whisper.cpp CMake configure step
cannot find nvcc:

CUDAToolkit_NVCC_EXECUTABLE:FILEPATH=CUDAToolkit_NVCC_EXECUTABLE-NOTFOUND

On systems where nvcc is found, CUDA 13.x / CCCL also requires C++17 for CUDA sources. The top-level project uses
C++17, but the nested whisper.cpp ExternalProject did not receive an explicit CUDA standard. This can fail later
during CUDA compilation with:

CUB requires at least C++17
libcu++ requires at least C++ 17
Thrust requires at least C++17

Why this follows upstream/toolchain behavior

CMake documents CUDA toolkit discovery in this order: the enabled CUDA compiler, CMAKE_CUDA_COMPILER / CUDACXX,
CUDAToolkit_ROOT, CUDA_PATH, then PATH, with platform defaults only after that. See:

• https://cmake.org/cmake/help/latest/module/FindCUDAToolkit.html
• https://cmake.org/cmake/help/latest/envvar/CUDACXX.html

CMake also documents CUDAHOSTCXX as the preferred host compiler input for CUDA compilation, which is cached as
CMAKE_CUDA_HOST_COMPILER:

• https://cmake.org/cmake/help/latest/envvar/CUDAHOSTCXX.html
• https://cmake.org/cmake/help/latest/variable/CMAKE_CUDA_HOST_COMPILER.html

NVIDIA's Linux installation guide says CUDA post-install setup must put the toolkit bin directory in PATH, and its
FAQ explicitly points to missing PATH setup when nvcc is not found:

• https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#post-installation-actions
• https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#why-do-i-see-nvcc-no-such-file-or-directory-when-i-try-to-build-a-cuda-application

The Arch cuda package matches that model: it installs the toolkit under /opt/cuda, ships /etc/profile.d/cuda.sh
with CUDA_PATH=/opt/cuda, adds /opt/cuda/bin to PATH, and sets NVCC_CCBIN=/usr/bin/g++-15.

• https://gitlab.archlinux.org/archlinux/packaging/packages/cuda/-/raw/main/cuda.sh
• https://gitlab.archlinux.org/archlinux/packaging/packages/cuda/-/blob/main/PKGBUILD

Why this is not only a local setup issue

The current package logic enables CUDA based only on package presence. That works only if the build environment already
has the CUDA profile loaded. A clean AUR helper / makepkg environment can have cuda installed but still not expose
nvcc through PATH.

This fix keeps user-configured CUDA locations working (CUDACXX, CUDAToolkit_ROOT, CUDA_PATH, CUDA_HOME, PATH)
and uses /opt/cuda only as the Arch-package fallback. It does not require all systems to install CUDA in /opt/cuda.

Verification

On Arch/CachyOS with:

• cuda 13.3.1
• gcc15
• cmake 4.3.4

I verified a clean CMake build of the CUDA ExternalProject path:

cmake -S <source-dir> -B <build-dir> \
  -DWITH_DESKTOP=ON \
  -DWITH_SFOS=OFF \
  -DWITH_PY=OFF \
  -DWITH_TESTS=OFF \
  -DDOWNLOAD_LIBSTT=OFF \
  -DBUILD_OPENBLAS=OFF \
  -DBUILD_VOSK=OFF \
  -DBUILD_BERGAMOT=OFF \
  -DBUILD_RHVOICE=OFF \
  -DBUILD_RHVOICE_MODULE=OFF \
  -DBUILD_WHISPERCPP_VULKAN=OFF \
  -DBUILD_WHISPERCPP_CUBLAS=ON \
  -DCUDAToolkit_ROOT=/opt/cuda \
  -DCMAKE_CUDA_COMPILER=/opt/cuda/bin/nvcc \
  -DCMAKE_CUDA_HOST_COMPILER=/usr/bin/g++-15 \
  -DCMAKE_CUDA_ARCHITECTURES=native

cmake --build <build-dir> --target whispercppcublas -- -j2

The generated nested configure command includes:

-DCUDAToolkit_ROOT=/opt/cuda
-DCMAKE_CUDA_COMPILER=/opt/cuda/bin/nvcc
-DCMAKE_CUDA_HOST_COMPILER=/usr/bin/g++-15
-DCMAKE_CUDA_STANDARD=17
-DCMAKE_CUDA_STANDARD_REQUIRED=ON

The build completed successfully:

[100%] Built target whispercppcublas