# Maintainer: opencode (test build)
# Contributor: Bink (original design)
# Difference from upstream: blobless partial clone + shared persistent source
# cache (see ggml-cuda-git). Full history (~418 MiB) is never downloaded;
# commits/tags are kept so `git describe` still works.

pkgname=llama.cpp-cuda-git
_pkgname="${pkgname%-cuda-git}"
_srccache="${LLAMACPP_SRC_CACHE:-${XDG_CACHE_HOME:-${HOME}/.cache}/llama.cpp-git}"
pkgver=b10268.r3.2f56fc34
pkgrel=1
pkgdesc="Port of Facebook's LLaMA model in C/C++ (with NVIDIA CUDA optimizations)"
arch=(x86_64 aarch64)
url='https://github.com/ggml-org/llama.cpp'
license=('MIT')
backup=('etc/conf.d/llama.cpp')
depends=(
  ggml-cuda-git
  curl
  gcc-libs
  glibc
  openssl
)
makedepends=(
  cmake
  cuda
  git
  ninja
)
optdepends=(
  'ccache: greatly reduce package re-build time'
  'nccl: needed for multi-GPU parallelism'
  'python-numpy: needed for convert_hf_to_gguf.py'
  'python-safetensors: needed for convert_hf_to_gguf.py'
  'python-sentencepiece: needed for convert_hf_to_gguf.py'
  'python-pytorch: needed for convert_hf_to_gguf.py'
  'python-transformers: needed for convert_hf_to_gguf.py'
)
provides=("${_pkgname}")
conflicts=("${_pkgname}")
source=(
  llama.cpp.conf
  llama.cpp.service
)
sha256sums=('53fa70cfe40cb8a3ca432590e4f76561df0f129a31b121c9b4b34af0da7c4d87'
            '0377d08a07bda056785981d3352ccd2dbc0387c4836f91fb73e6b790d836620d')

pkgver() {
  cd "${_pkgname}" || 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 (same one used by ggml-cuda-git).
  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}/${_pkgname}"
  git -C "${srcdir}/${_pkgname}" checkout --detach origin/master
}

build() {
  local _commit_id _build_number
  _commit_id=$(git -C "${_pkgname}" rev-parse HEAD)
  _build_number=$(git -C "${_pkgname}" rev-list --count HEAD)

  local _cmake_options=(
    -G Ninja
    -B build
    -S "${_pkgname}"
    -DCMAKE_BUILD_TYPE=Release
    -DCMAKE_INSTALL_PREFIX='/usr'
    -DBUILD_SHARED_LIBS=ON
    -DLLAMA_BUILD_TESTS=OFF
    -DLLAMA_USE_SYSTEM_GGML=ON
    -DLLAMA_BUILD_SERVER=ON
    -DLLAMA_BUILD_NUMBER="${_build_number}"
    -DLLAMA_BUILD_COMMIT="${_commit_id}"
    -DLLAMA_OPENSSL=ON
    -Wno-dev
  )

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

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

package() {
  DESTDIR="${pkgdir}" cmake --install build
  install -Dm644 "${_pkgname}/LICENSE" "${pkgdir}/usr/share/licenses/${pkgname}/LICENSE"
  install -Dm644 "llama.cpp.conf" "${pkgdir}/etc/conf.d/llama.cpp"
  install -Dm644 "llama.cpp.service" "${pkgdir}/usr/lib/systemd/system/llama.cpp.service"
}