Arch Linux User Repository

I've now updated this to depend on ggml-cuda-git (which I maintain) to provide CUDA optimised GGML, so hopefully no more issues with unreliable dependencies like the now deleted libggml-cuda-git.

I've now updated this to depend on ggml-cuda-git (which I maintain) to provide CUDA optimised GGML, so hopefully no more issues with unreliable dependencies like the now deleted libggml-cuda-git.

@graysky, the runtime dependency on cuda falls to the cuda optimised libggml package which is building against and depending on cuda, not this package.

For example, you'll need to add cuda as a dependency to libggml-cuda.

I think it's worth noting that if you've pre-installed a cuda optimised provider of libggml (eg: yours, libggml-cuda-git, or llama.cpp-cuda-git), installing the standard whisper.cpp package then results in cuda optimised whisper.cpp.

All this package seeks to do, is be explicit about that dependency.

I submitted libggml-cuda can you plz change the depends dropping the -git variant?

-depends=('ffmpeg' 'libggml-cuda-git' 'nvidia-utils' 'sdl2-compat')
+depends=('ffmpeg' 'libggml-cuda' 'nvidia-utils' 'sdl2-compat' 'cuda')

missing a build dep:

--- a/PKGBUILD
+++ b/PKGBUILD
@@ -9,7 +9,7 @@ pkgdesc="High-performance inference of OpenAI's Whisper automatic speech recogni
 arch=('aarch64' 'x86_64')
 url="https://github.com/ggml-org/whisper.cpp"
 license=("MIT")
-depends=('ffmpeg' 'libggml-cuda-git' 'nvidia-utils' 'sdl2-compat')
+depends=('ffmpeg' 'libggml-cuda-git' 'nvidia-utils' 'sdl2-compat' 'cuda')
 optdepends=(
   'openblas: CPU-accelerated matrix ops (only if system ggml was built with BLAS)'
 )
@@ -17,6 +17,7 @@ conflicts=("${_pkgbase}")
 provides=("${_pkgbase}")
 makedepends=(
   'cmake'
+  'cuda'
   'ninja'
 )
 source=(

@graysky, this is using CUDA as provided by dependencies, namely libggml-cuda-git or llama.cpp.cuda-git, not by itself, so it remains compatible with other tools a user might be using, and the user is not needing to compile again something that already exists on the system.

So yes, it is actually building the CUDA flavour, but in a way that plays nice with other installed packages.

Note the flag you've enabled:

build -DGGML_CUDA=1

And note what is already explicitly specified as a dependency:

libggml-cuda-git

So you can force this package to build GGML_CUDA, but it is redundant and may even conflict, as it will also be building that as part of libggml-cuda-git or llama.cpp-cuda-git.

Here's a functional PKGBUILD: https://gist.github.com/graysky2/022b39a65b4dbdcf97f8ed9eb2e29b67

Are you sure this is actually building the cuda flavor? For one, you are missing a couple of build dependencies, including CUDA and GCC15. for two, you are not building with the CUDA flags.

From: https://github.com/ggml-org/whisper.cpp#nvidia-gpu-support

With NVIDIA cards the processing of the models is done efficiently on the GPU via cuBLAS and custom CUDA kernels.
First, make sure you have installed cuda: https://developer.nvidia.com/cuda-downloads

Now build whisper.cpp with CUDA support:

cmake -B build -DGGML_CUDA=1
cmake --build build -j --config Release
or for newer NVIDIA GPU's (RTX 5000 series):

cmake -B build -DGGML_CUDA=1 -DCMAKE_CUDA_ARCHITECTURES="86"
cmake --build build -j --config Release

@xubaiw, you might consider using llama.cpp-cuda-git instead. I've specifically configured that to satisfy the optimised-for-CUDA libggml-cuda-git dependency for this package.

:: Downloading PKGBUILDs...
 PKGBUILDs up to date
 nothing new to review
fetching devel info...
==> Making package: libggml-cuda-git 0.9.4.r387.g3e9f2ba-1 (Сб 25 апр 2026 10:27:44)
==> Retrieving sources...
  -> Updating libggml git repo...
==> Validating source files with sha256sums...
    libggml ... Skipped
==> Making package: libggml-cuda-git 0.9.4.r387.g3e9f2ba-1 (Сб 25 апр 2026 10:27:45)
==> Checking runtime dependencies...
==> Checking buildtime dependencies...
==> Retrieving sources...
  -> Updating libggml git repo...
==> Validating source files with sha256sums...
    libggml ... Skipped
==> Removing existing $srcdir/ directory...
==> Extracting sources...
  -> Creating working copy of libggml git repo...
Cloning into 'libggml'...
done.
==> Starting prepare()...
==> Starting pkgver()...
==> Updated version: libggml-cuda-git 0.10.0.r1.g8be60f8-1
==> Sources are ready.
libggml-cuda-git-0.9.4.r387.g3e9f2ba-1: parsing pkg list...
==> Making package: libggml-cuda-git 0.10.0.r1.g8be60f8-1 (Сб 25 апр 2026 10:27:46)
==> Checking runtime dependencies...
==> Checking buildtime dependencies...
==> WARNING: Using existing $srcdir/ tree
==> Starting pkgver()...
==> Removing existing $pkgdir/ directory...
==> Starting build()...
Building with OpenBLAS + Vulkan + x86-64 support...
-- The C compiler identification is GNU 15.2.1
-- The CXX compiler identification is GNU 15.2.1
-- The ASM compiler identification is GNU
-- Found assembler: /usr/bin/cc
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working C compiler: /usr/bin/cc - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Setting GGML_NATIVE_DEFAULT to OFF
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD - Success
-- Found Threads: TRUE
-- Warning: ccache not found - consider installing it for faster compilation or disable this warning with GGML_CCACHE=OFF
-- CMAKE_SYSTEM_PROCESSOR: x86_64
-- GGML_SYSTEM_ARCH: x86
-- Including CPU backend
-- Found OpenMP_C: -fopenmp (found version "4.5")
-- Found OpenMP_CXX: -fopenmp (found version "4.5")
-- Found OpenMP: TRUE (found version "4.5")
-- x86 detected
-- Adding CPU backend variant ggml-cpu:
-- Looking for sgemm_
-- Looking for sgemm_ - found
-- Found BLAS: /usr/lib/libopenblas.so
-- BLAS found, Libraries: /usr/lib/libopenblas.so
-- Found PkgConfig: /usr/bin/pkg-config (found version "2.5.1")
-- Checking for module 'openblas64'
--   Package 'openblas64' not found
-- Checking for module 'openblas'
--   Found openblas, version 0.3.32
-- BLAS found, Includes: /usr/include/openblas
-- Including BLAS backend
-- Could not find `nvcc` executable in any searched paths, please set CUDAToolkit_ROOT
CMake Error at src/ggml-cuda/CMakeLists.txt:267 (message):
  CUDA Toolkit not found


-- Configuring incomplete, errors occurred!
==> ERROR: A failure occurred in build().
    Aborting...
error: failed to build 'libggml-cuda-git-0.9.4.r387.g3e9f2ba-1':
error: can't build whisper.cpp-cuda-1.8.4-1, deps not satisfied: libggml-cuda-git
error: packages failed to build: libggml-cuda-git-0.9.4.r387.g3e9f2ba-1  whisper.cpp-cuda-1.8.4-1