# Maintainer: opencode (test build)
# Contributor: Bink (original design)
# Difference from upstream: blobless partial clone + shared persistent source
# cache, so the ~418 MiB full history is never downloaded. Commits and tags
# are kept, so `git describe` still works.

pkgname=ggml-cuda-git
_srcname=llama.cpp
_srccache="${LLAMACPP_SRC_CACHE:-${XDG_CACHE_HOME:-${HOME}/.cache}/llama.cpp-git}"
pkgver=b10268.r3.2f56fc34
pkgrel=1
epoch=1
pkgdesc="Tensor library for machine learning (with NVIDIA CUDA optimizations)"
arch=(x86_64 aarch64)
url='https://github.com/ggml-org/ggml'
license=('MIT')
: "${aur_ggml_build_universal:=false}"
depends=(
  cuda
  gcc-libs
  glibc
  nvidia-utils
)
makedepends=(
  cmake
  gcc15   # CUDA does not yet support GCC 16
  git
  ninja
)
optdepends=(
  'nccl: needed for multi-GPU parallelism (rebuild required)'
  'rdma-core: RDMA transport for RPC backend (rebuild required)'
)
provides=(
  ggml-cuda-git
  libggml
  libggml.so
  ggml
)
conflicts=(
  libggml
  ggml
)
# No git source here: the clone happens in prepare() from a shared blobless
# cache, so ggml-cuda-git and llama.cpp-cuda-git never download the history twice.
source=()
sha256sums=()

pkgver() {
  cd "${_srcname}" || return 1
  if git describe --long --tags >/dev/null 2>&1; then
    git describe --long --tags | sed 's/\([^-]*-\)g/r\1/;s/-/./g'
  else
    printf 'r%s.%s' "$(git rev-list --count HEAD)" "$(git rev-parse --short HEAD)"
  fi
}

prepare() {
  # Shared blobless source cache (dedup across packages + persistent across builds).
  if [[ ! -d "${_srccache}/.git" ]]; then
    git clone --filter=blob:none --no-checkout --single-branch \
      https://github.com/ggml-org/llama.cpp.git "${_srccache}"
  else
    git -C "${_srccache}" fetch --prune origin
  fi
  git -C "${_srccache}" checkout --force --detach origin/master
  # Clone from the shared cache (copy across filesystems). The network download
  # happens only once, into the cache; this is fast local disk I/O.
  git clone --local --no-hardlinks --no-checkout "${_srccache}" "${srcdir}/${_srcname}"
  git -C "${srcdir}/${_srcname}" checkout --detach origin/master
}

build() {
  if ! type -P nvcc &>/dev/null && [[ -d /opt/cuda/bin ]]; then
    export PATH="/opt/cuda/bin:$PATH"
  fi

  # Wrap ggml/ in a tiny standalone project; see upstream comment:
  # building ggml/ directly trips llama.cpp's "standalone" detection.
  local _wrapper="${srcdir}/_ggml_wrapper"
  mkdir -p "${_wrapper}"
  cat > "${_wrapper}/CMakeLists.txt" <<EOF
cmake_minimum_required(VERSION 3.14)
project(ggml-cuda-git C CXX)
add_subdirectory("${srcdir}/${_srcname}/ggml" ggml)
EOF

  local _nvcc_host_cxx="${CUDAHOSTCXX:-/usr/bin/g++-15}"

  local _cmake_options=(
    -G Ninja
    -B build
    -S "${_wrapper}"
    -DCMAKE_BUILD_TYPE=Release
    -DCMAKE_INSTALL_PREFIX='/usr'
    -DCMAKE_CUDA_HOST_COMPILER="${_nvcc_host_cxx}"
    -DBUILD_SHARED_LIBS=ON
    -DGGML_ALL_WARNINGS=OFF
    -DGGML_ALL_WARNINGS_3RD_PARTY=OFF
    -DGGML_BUILD_EXAMPLES=OFF
    -DGGML_BUILD_TESTS=OFF
    -DGGML_OPENMP=ON
    -DGGML_LTO=ON
    -DGGML_RPC=ON
    -DGGML_CUDA=ON
    -DGGML_CUDA_FA_ALL_QUANTS=ON
    -DGGML_CUDA_COMPRESSION_MODE=speed
    -DGGML_CUDA_GRAPHS=ON
    -DGGML_LLAMAFILE=ON
    -DGGML_BLAS=OFF
    -DGGML_VULKAN=OFF
    -Wno-dev
  )

  if [[ ${aur_ggml_build_universal} == true ]]; then
    echo "Building universal binary [aur_ggml_build_universal == true]"
    _cmake_options+=(
      -DGGML_BACKEND_DL=ON
      -DGGML_NATIVE=OFF
      -DGGML_CPU_ALL_VARIANTS=ON
    )
  else
    _cmake_options+=(
      -DGGML_BACKEND_DL=OFF
      -DGGML_NATIVE=ON
      -DCMAKE_CUDA_ARCHITECTURES=native
    )
  fi

  if [[ -n "${aur_ggml_cmakeopts:-}" ]]; then
    echo "Applying custom CMake options: ${aur_ggml_cmakeopts}"
    _cmake_options+=(${aur_ggml_cmakeopts})
  fi

  cmake "${_cmake_options[@]}"
  cmake --build build
}

package() {
  DESTDIR="${pkgdir}" cmake --install build
  install -Dm644 "${_srcname}/LICENSE" "${pkgdir}/usr/share/licenses/${pkgname}/LICENSE"
}