ilyar · March 5, 2026 04:50
diff --git a/install-llama-cpp.sh b/install-llama-cpp.sh
 #!/usr/bin/env bash

 # Install llama.cpp prebuilt binaries from GitHub releases https://github.com/ggml-org/llama.cpp/releases
 #
 # curl -fsSL <URL_RAW> | bash
 #
 # Supports:
 # - macOS (arm64/x64)
 # - Linux Ubuntu (x64/s390x, cpu/vulkan/rocm7.2)
 # - Windows (cpu/cuda12.4/cuda13.1/vulkan/sycl/hip/opencl-adreno)
 # - openEuler (310p/910b + optional aclgraph)
 #
 # Env vars:
 #   LLAMA_TAG               e.g. b8201 (default: latest)
 #   LLAMA_BACKEND           auto|cpu|vulkan|rocm|cuda12|cuda13|sycl|hip|opencl-adreno|xcframework (default: auto)
 #   LLAMA_PREFIX            install prefix (default: /usr/local if writable or sudo available; else ~/.local)
 #   LLAMA_NO_SUDO           set to 1 to never use sudo
 #   LLAMA_ENABLE_SYSTEMD    set to 1 to install + start systemd service on Linux (default: 0)
 #   LLAMA_SERVER_HOST       default 127.0.0.1
 #   LLAMA_SERVER_PORT       default 8080
 #   LLAMA_SERVER_ARGS       extra args for llama-server
 #   LLAMA_OE_FAMILY         openEuler only: 310p|910b (default: 310p)
 #   LLAMA_OE_ACLGRAPH       openEuler only: 0|1 (default: 0)
 #
 # Examples:
 #   curl -fsSL <URL_RAW>/install-llama-cpp.sh | bash
 #   LLAMA_BACKEND=vulkan curl -fsSL <URL_RAW>/install-llama-cpp.sh | bash
 #   LLAMA_BACKEND=auto LLAMA_ENABLE_SYSTEMD=1 curl -fsSL <URL_RAW>/install-llama-cpp.sh | bash
 #   LLAMA_PREFIX=~/.local curl -fsSL <URL_RAW>/install-llama-cpp.sh | bash
 #   LLAMA_BACKEND=cuda12 curl -fsSL <URL_RAW>/install-llama-cpp.sh | bash # Windows (Git Bash)

 main() {
  set -euo pipefail

  red="$( (tput bold || :; tput setaf 1 || :) 2>/dev/null )"
  plain="$( (tput sgr0 || :) 2>/dev/null )"

  status() { echo ">>> $*" >&2; }
  warning() { echo "${red}WARNING:${plain} $*" >&2; }
  error() { echo "${red}ERROR:${plain} $*" >&2; exit 1; }

  available() { command -v "$1" >/dev/null 2>&1; }

  require() {
    local missing=""
    for tool in "$@"; do
      if ! available "$tool"; then
        missing="$missing $tool"
      fi
    done
    echo "$missing"
  }

  TMP_DIR="$(mktemp -d)"
  cleanup() { rm -rf "$TMP_DIR"; }
  trap cleanup EXIT

  # ---------- config ----------
  REPO="ggml-org/llama.cpp"
  API_BASE="https://api.github.com/repos/${REPO}/releases"
  UA="llama.cpp-prebuilt-installer/2.0"

  TAG="${LLAMA_TAG:-"$(curl -fsSL https://api.github.com/repos/ggml-org/llama.cpp/releases/latest | sed -n 's/.*"tag_name": *"\([^"]*\)".*/\1/p')"}"
  BACKEND="${LLAMA_BACKEND:-auto}"

  ENABLE_SYSTEMD="${LLAMA_ENABLE_SYSTEMD:-0}"
  SERVER_HOST="${LLAMA_SERVER_HOST:-127.0.0.1}"
  SERVER_PORT="${LLAMA_SERVER_PORT:-8080}"
  SERVER_ARGS="${LLAMA_SERVER_ARGS:-}"

  OE_FAMILY="${LLAMA_OE_FAMILY:-310p}"     # openEuler: 310p or 910b
  OE_ACLGRAPH="${LLAMA_OE_ACLGRAPH:-0}"    # openEuler: 1 => -aclgraph

  # ---------- deps ----------
  NEEDS="$(require curl uname grep cut sed awk)"
  if [ -n "$NEEDS" ]; then
    status "Missing required tools:"
    for n in $NEEDS; do echo "  - $n"; done
    exit 1
  fi

  # ---------- sudo policy ----------
  SUDO=""
  if [ "${LLAMA_NO_SUDO:-0}" = "1" ]; then
    SUDO=""
  else
    if [ "$(id -u)" -ne 0 ]; then
      if available sudo; then
        SUDO="sudo"
      fi
    fi
  fi

  # ---------- detect OS / arch ----------
  OS="$(uname -s)"
  ARCH_RAW="$(uname -m)"

  case "$ARCH_RAW" in
    x86_64|amd64) ARCH="x64" ;;
    aarch64|arm64) ARCH="arm64" ;;
    s390x) ARCH="s390x" ;;
    *) error "Unsupported architecture: $ARCH_RAW" ;;
  esac

  IS_WINDOWS=0
  IS_MAC=0
  IS_LINUX=0

  case "$OS" in
    Darwin) IS_MAC=1 ;;
    Linux) IS_LINUX=1 ;;
    MINGW*|MSYS*|CYGWIN*) IS_WINDOWS=1 ;;
    *) error "Unsupported OS: $OS" ;;
  esac

  # openEuler is Linux, detect via /etc/os-release
  IS_OPENEULER=0
  if [ "$IS_LINUX" = "1" ] && [ -r /etc/os-release ]; then
    if grep -qi '^ID=.*openeuler' /etc/os-release; then
      IS_OPENEULER=1
    fi
  fi

  # ---------- choose prefix ----------
  if [ -n "${LLAMA_PREFIX:-}" ]; then
    PREFIX="$LLAMA_PREFIX"
  else
    # If /usr/local is writable OR we have sudo, default to /usr/local, else ~/.local
    if [ -w /usr/local ] || [ -n "$SUDO" ]; then
      PREFIX="/usr/local"
    else
      PREFIX="${HOME}/.local"
    fi
  fi

  BIN_DIR="${PREFIX}/bin"
  ROOT_DIR="${PREFIX}/lib/llama.cpp"  # keep versioned installs here
  mkdir -p "$TMP_DIR"

  # ---------- fetch release tag ----------
  fetch_release_json() {
    local url="$1"
    local out="$2"
    curl -fsSL -H "Accept: application/vnd.github+json" -H "User-Agent: ${UA}" "$url" > "$out" \
      || error "Failed to fetch release metadata (GitHub API rate limit? try later)"
  }

  RELEASE_JSON="$TMP_DIR/release.json"
  if [ -n "$TAG" ]; then
    status "Fetching release info for tag: $TAG"
    fetch_release_json "${API_BASE}/tags/${TAG}" "$RELEASE_JSON"
  else
    status "Fetching latest release info..."
    fetch_release_json "${API_BASE}/latest" "$RELEASE_JSON"
    TAG="$(grep -m1 '"tag_name"' "$RELEASE_JSON" | cut -d '"' -f4 || true)"
    [ -n "$TAG" ] || error "Could not parse tag_name from GitHub API response"
  fi

  status "Selected release: $TAG"

  # ---------- auto backend detection ----------
  # Important: Linux prebuilt list here has: cpu/vulkan/rocm7.2 only (no CUDA for Linux in your list)
  # Windows has cuda12/cuda13 variants + cudart
  auto_backend() {
    if [ "$IS_MAC" = "1" ]; then
      # No separate metal variant in your list; prebuilt is "macos-ARCH"
      echo "cpu"
      return 0
    fi

    if [ "$IS_WINDOWS" = "1" ]; then
      # Prefer CUDA if nvidia-smi exists, else vulkan if vulkaninfo exists, else cpu
      if available nvidia-smi; then
        # Pick cuda12 by default (more common driver availability); user can override
        echo "cuda12"
      elif available vulkaninfo; then
        echo "vulkan"
      else
        echo "cpu"
      fi
      return 0
    fi

    # Linux (including openEuler): detect ROCm, then Vulkan, else CPU
    if available rocminfo; then
      echo "rocm"
    elif available vulkaninfo; then
      echo "vulkan"
    else
      echo "cpu"
    fi
  }

  if [ "$BACKEND" = "auto" ]; then
    BACKEND="$(auto_backend)"
    status "Auto-selected backend: $BACKEND"
  else
    status "Requested backend: $BACKEND"
  fi

  # ---------- build asset filename ----------
  ASSET=""
  ASSET2="" # extra (cudart) for some Windows CUDA builds
  EXT="tar.gz"

  if [ "$IS_MAC" = "1" ]; then
    case "$BACKEND" in
      cpu|vulkan|rocm) : ;; # tolerate; mac uses same tarball
      xcframework)
        # iOS XCFramework
        EXT="zip"
        ASSET="llama-${TAG}-xcframework.zip"
        ;;
      *)
        warning "macOS: backend '$BACKEND' ignored; using macos prebuilt"
        ;;
    esac
    if [ -z "$ASSET" ]; then
      ASSET="llama-${TAG}-bin-macos-${ARCH}.tar.gz"
      EXT="tar.gz"
    fi

  elif [ "$IS_WINDOWS" = "1" ]; then
    EXT="zip"
    case "$BACKEND" in
      cpu)
        ASSET="llama-${TAG}-bin-win-cpu-${ARCH}.zip"
        ;;
      cuda12)
        [ "$ARCH" = "x64" ] || error "Windows CUDA builds are x64 only in your list"
        ASSET="llama-${TAG}-bin-win-cuda-12.4-x64.zip"
        ASSET2="cudart-llama-bin-win-cuda-12.4-x64.zip"
        ;;
      cuda13)
        [ "$ARCH" = "x64" ] || error "Windows CUDA builds are x64 only in your list"
        ASSET="llama-${TAG}-bin-win-cuda-13.1-x64.zip"
        # Your list shows only cudart for 12.4; keep 13 without extra unless they add it later
        ;;
      vulkan)
        [ "$ARCH" = "x64" ] || error "Windows Vulkan build in your list is x64 only"
        ASSET="llama-${TAG}-bin-win-vulkan-x64.zip"
        ;;
      sycl)
        [ "$ARCH" = "x64" ] || error "Windows SYCL build in your list is x64 only"
        ASSET="llama-${TAG}-bin-win-sycl-x64.zip"
        ;;
      hip)
        [ "$ARCH" = "x64" ] || error "Windows HIP Radeon build in your list is x64 only"
        ASSET="llama-${TAG}-bin-win-hip-radeon-x64.zip"
        ;;
      opencl-adreno)
        [ "$ARCH" = "arm64" ] || error "Windows OpenCL Adreno build in your list is arm64 only"
        ASSET="llama-${TAG}-bin-win-opencl-adreno-arm64.zip"
        ;;
      *)
        error "Unsupported Windows backend: $BACKEND"
        ;;
    esac

  else
    # Linux / openEuler
    if [ "$IS_OPENEULER" = "1" ]; then
      # openEuler assets are named differently and don't include x64/arm64 tags in the same style
      # Map ARCH -> x86/aarch64 in names
      OE_ARCH=""
      case "$ARCH" in
        x64) OE_ARCH="x86" ;;
        arm64) OE_ARCH="aarch64" ;;
        *) error "openEuler assets in your list support x86 and aarch64 only; got: $ARCH" ;;
      esac

      case "$OE_FAMILY" in
        310p|910b) : ;;
        *) error "openEuler: LLAMA_OE_FAMILY must be 310p or 910b" ;;
      esac

      if [ "$OE_ACLGRAPH" = "1" ]; then
        ASSET="llama-${TAG}-bin-${OE_FAMILY}-openEuler-${OE_ARCH}-aclgraph.tar.gz"
      else
        ASSET="llama-${TAG}-bin-${OE_FAMILY}-openEuler-${OE_ARCH}.tar.gz"
      fi
      EXT="tar.gz"

      # Note: backend selection is ignored for openEuler because your assets are family-based
      if [ "$BACKEND" != "cpu" ]; then
        status "openEuler: ignoring backend='$BACKEND' (assets are 310p/910b based)"
      fi

    else
      # Ubuntu assets (Linux)
      case "$BACKEND" in
        cpu)
          if [ "$ARCH" = "s390x" ]; then
            ASSET="llama-${TAG}-bin-ubuntu-s390x.tar.gz"
          else
            [ "$ARCH" = "x64" ] || error "Ubuntu CPU assets in your list are x64 and s390x only"
            ASSET="llama-${TAG}-bin-ubuntu-x64.tar.gz"
          fi
          ;;
        vulkan)
          [ "$ARCH" = "x64" ] || error "Ubuntu Vulkan asset in your list is x64 only"
          ASSET="llama-${TAG}-bin-ubuntu-vulkan-x64.tar.gz"
          ;;
        rocm)
          [ "$ARCH" = "x64" ] || error "Ubuntu ROCm asset in your list is x64 only"
          ASSET="llama-${TAG}-bin-ubuntu-rocm-7.2-x64.tar.gz"
          ;;
        *)
          error "Unsupported Linux backend: $BACKEND (use cpu|vulkan|rocm|auto)"
          ;;
      esac
      EXT="tar.gz"
    fi
  fi

  [ -n "$ASSET" ] || error "Internal: failed to choose asset"

  # ---------- download ----------
  dl() {
    local filename="$1"
    local dest="$2"
    local url="https://github.com/${REPO}/releases/download/${TAG}/${filename}"
    status "Downloading: $filename"
    curl --fail --show-error --location --progress-bar -H "User-Agent: ${UA}" \
      -o "$dest" "$url" || error "Download failed: $filename"
  }

  ARCHIVE1="$TMP_DIR/asset1.${EXT}"
  dl "$ASSET" "$ARCHIVE1"

  ARCHIVE2=""
  if [ -n "$ASSET2" ]; then
    ARCHIVE2="$TMP_DIR/asset2.zip"
    dl "$ASSET2" "$ARCHIVE2"
  fi

  # ---------- extract ----------
  EXTRACT_DIR="$TMP_DIR/extract"
  mkdir -p "$EXTRACT_DIR"

  if [ "$EXT" = "zip" ]; then
    NEEDS="$(require unzip)"
    [ -z "$NEEDS" ] || error "Missing required tool: unzip"
    unzip -q "$ARCHIVE1" -d "$EXTRACT_DIR"
  else
    NEEDS="$(require tar)"
    [ -z "$NEEDS" ] || error "Missing required tool: tar"
    tar -xzf "$ARCHIVE1" -C "$EXTRACT_DIR"
  fi

  if [ -n "$ARCHIVE2" ]; then
    NEEDS="$(require unzip)"
    [ -z "$NEEDS" ] || error "Missing required tool: unzip"
    unzip -q "$ARCHIVE2" -d "$EXTRACT_DIR"
  fi

  # ---------- install layout ----------
  # We keep a versioned directory under PREFIX/lib/llama.cpp and then link executables into PREFIX/bin
  INSTALL_ID="${TAG}-$(uname -s | tr '[:upper:]' '[:lower:]')-${ARCH}-${BACKEND}"
  INSTALL_DIR="${ROOT_DIR}/${INSTALL_ID}"

  status "Installing into: $INSTALL_DIR"
  if [ -n "$SUDO" ] && [ "$PREFIX" = "/usr/local" ]; then
    $SUDO mkdir -p "$BIN_DIR" "$ROOT_DIR"
    $SUDO rm -rf "$INSTALL_DIR"
    $SUDO mkdir -p "$INSTALL_DIR"
    $SUDO cp -R "$EXTRACT_DIR"/. "$INSTALL_DIR"/
  else
    mkdir -p "$BIN_DIR" "$ROOT_DIR"
    rm -rf "$INSTALL_DIR"
    mkdir -p "$INSTALL_DIR"
    cp -R "$EXTRACT_DIR"/. "$INSTALL_DIR"/
  fi

  # ---------- link executables ----------
  link_one() {
    local src="$1"
    local name="$2"
    if [ -n "$SUDO" ] && [ "$PREFIX" = "/usr/local" ]; then
      $SUDO ln -sf "$src" "${BIN_DIR}/${name}" 2>/dev/null || $SUDO cp -f "$src" "${BIN_DIR}/${name}"
    else
      ln -sf "$src" "${BIN_DIR}/${name}" 2>/dev/null || cp -f "$src" "${BIN_DIR}/${name}"
    fi
  }

  status "Linking binaries into: $BIN_DIR"

  if [ "$IS_WINDOWS" = "1" ]; then
    # link *.exe and *.bat
    wins=()
    while IFS= read -r line; do wins+=("$line"); done < <(find "$INSTALL_DIR" -type f \( -iname '*.exe' -o -iname '*.bat' \) 2>/dev/null || true)
    [ "${#wins[@]}" -gt 0 ] || warning "No .exe/.bat found. Asset layout may have changed"
    for f in "${wins[@]}"; do
      link_one "$f" "$(basename "$f")"
    done
  else
    # link all executable files (best effort)
    nix=()
    while IFS= read -r line; do nix+=("$line"); done < <(find "$INSTALL_DIR" -type f -perm -111 2>/dev/null || true)
    [ "${#nix[@]}" -gt 0 ] || warning "No executable files found. Asset layout may have changed"
    for f in "${nix[@]}"; do
      link_one "$f" "$(basename "$f")"
    done
    # friendly alias
    if [ -f "${BIN_DIR}/llama-cli" ]; then
      link_one "${BIN_DIR}/llama-cli" "llama"
    fi
  fi

  # ---------- optional systemd (Linux only) ----------
  configure_systemd() {
    [ "$IS_LINUX" = "1" ] || return 0
    [ "$ENABLE_SYSTEMD" = "1" ] || return 0
    available systemctl || { warning "systemctl not found; skipping systemd service"; return 0; }

    local server_path="${BIN_DIR}/llama-server"
    [ -x "$server_path" ] || { warning "llama-server not found in ${BIN_DIR}; skipping systemd"; return 0; }

    if [ -z "$SUDO" ]; then
      warning "No sudo available; cannot install system-wide systemd service. (Set LLAMA_PREFIX=~/.local and use user systemd manually.)"
      return 0
    fi

    status "Creating llama.cpp systemd service (llama-cpp.service)..."
    $SUDO tee /etc/systemd/system/llama-cpp.service >/dev/null <<EOF
 [Unit]
 Description=llama.cpp server
 After=network-online.target
 Wants=network-online.target

 [Service]
 Type=simple
 ExecStart=${server_path} --host ${SERVER_HOST} --port ${SERVER_PORT} ${SERVER_ARGS}
 Restart=always
 RestartSec=2
 Environment="PATH=/usr/local/bin:/usr/bin:/bin:${BIN_DIR}"

 [Install]
 WantedBy=multi-user.target
 EOF

    status "Enabling and starting llama-cpp.service..."
    $SUDO systemctl daemon-reload
    $SUDO systemctl enable llama-cpp.service
    $SUDO systemctl restart llama-cpp.service

    status "llama.cpp server should be listening on ${SERVER_HOST}:${SERVER_PORT}"
  }

  configure_systemd

  status "Install complete"
  echo
  echo "Installed:   $INSTALL_DIR"
  echo "Binaries:    $BIN_DIR"
  echo
  if [ "$IS_WINDOWS" = "1" ]; then
    echo "Try (Git Bash):"
    echo "  ${BIN_DIR}/llama-cli.exe --help"
    echo "  ${BIN_DIR}/llama-server.exe --help"
  else
    echo "Try:"
    echo "  ${BIN_DIR}/llama-cli --help"
    echo "  ${BIN_DIR}/llama-server --help"
    echo
    if [ "$ENABLE_SYSTEMD" = "1" ] && [ "$IS_LINUX" = "1" ] && available systemctl; then
      echo "Service:"
      echo "  systemctl status llama-cpp.service"
    fi
  fi
 }

 main "$@"
	#!/usr/bin/env bash

	# Install llama.cpp prebuilt binaries from GitHub releases https://github.com/ggml-org/llama.cpp/releases
	#
	# curl -fsSL <URL_RAW> \| bash
	#
	# Supports:
	# - macOS (arm64/x64)
	# - Linux Ubuntu (x64/s390x, cpu/vulkan/rocm7.2)
	# - Windows (cpu/cuda12.4/cuda13.1/vulkan/sycl/hip/opencl-adreno)
	# - openEuler (310p/910b + optional aclgraph)
	#
	# Env vars:
	# LLAMA_TAG e.g. b8201 (default: latest)
	# LLAMA_BACKEND auto\|cpu\|vulkan\|rocm\|cuda12\|cuda13\|sycl\|hip\|opencl-adreno\|xcframework (default: auto)
	# LLAMA_PREFIX install prefix (default: /usr/local if writable or sudo available; else ~/.local)
	# LLAMA_NO_SUDO set to 1 to never use sudo
	# LLAMA_ENABLE_SYSTEMD set to 1 to install + start systemd service on Linux (default: 0)
	# LLAMA_SERVER_HOST default 127.0.0.1
	# LLAMA_SERVER_PORT default 8080
	# LLAMA_SERVER_ARGS extra args for llama-server
	# LLAMA_OE_FAMILY openEuler only: 310p\|910b (default: 310p)
	# LLAMA_OE_ACLGRAPH openEuler only: 0\|1 (default: 0)
	#
	# Examples:
	# curl -fsSL <URL_RAW>/install-llama-cpp.sh \| bash
	# LLAMA_BACKEND=vulkan curl -fsSL <URL_RAW>/install-llama-cpp.sh \| bash
	# LLAMA_BACKEND=auto LLAMA_ENABLE_SYSTEMD=1 curl -fsSL <URL_RAW>/install-llama-cpp.sh \| bash
	# LLAMA_PREFIX=~/.local curl -fsSL <URL_RAW>/install-llama-cpp.sh \| bash
	# LLAMA_BACKEND=cuda12 curl -fsSL <URL_RAW>/install-llama-cpp.sh \| bash # Windows (Git Bash)

	main() {
	set -euo pipefail

	red="$( (tput bold \|\| :; tput setaf 1 \|\| :) 2>/dev/null )"
	plain="$( (tput sgr0 \|\| :) 2>/dev/null )"

	status() { echo ">>> $*" >&2; }
	warning() { echo "${red}WARNING:${plain} $*" >&2; }
	error() { echo "${red}ERROR:${plain} $*" >&2; exit 1; }

	available() { command -v "$1" >/dev/null 2>&1; }

	require() {
	local missing=""
	for tool in "$@"; do
	if ! available "$tool"; then
	missing="$missing $tool"
	fi
	done
	echo "$missing"
	}

	TMP_DIR="$(mktemp -d)"
	cleanup() { rm -rf "$TMP_DIR"; }
	trap cleanup EXIT

	# ---------- config ----------
	REPO="ggml-org/llama.cpp"
	API_BASE="https://api.github.com/repos/${REPO}/releases"
	UA="llama.cpp-prebuilt-installer/2.0"

	TAG="${LLAMA_TAG:-"$(curl -fsSL https://api.github.com/repos/ggml-org/llama.cpp/releases/latest \| sed -n 's/."tag_name": "\([^"]\)"./\1/p')"}"
	BACKEND="${LLAMA_BACKEND:-auto}"

	ENABLE_SYSTEMD="${LLAMA_ENABLE_SYSTEMD:-0}"
	SERVER_HOST="${LLAMA_SERVER_HOST:-127.0.0.1}"
	SERVER_PORT="${LLAMA_SERVER_PORT:-8080}"
	SERVER_ARGS="${LLAMA_SERVER_ARGS:-}"

	OE_FAMILY="${LLAMA_OE_FAMILY:-310p}" # openEuler: 310p or 910b
	OE_ACLGRAPH="${LLAMA_OE_ACLGRAPH:-0}" # openEuler: 1 => -aclgraph

	# ---------- deps ----------
	NEEDS="$(require curl uname grep cut sed awk)"
	if [ -n "$NEEDS" ]; then
	status "Missing required tools:"
	for n in $NEEDS; do echo " - $n"; done
	exit 1
	fi

	# ---------- sudo policy ----------
	SUDO=""
	if [ "${LLAMA_NO_SUDO:-0}" = "1" ]; then
	SUDO=""
	else
	if [ "$(id -u)" -ne 0 ]; then
	if available sudo; then
	SUDO="sudo"
	fi
	fi
	fi

	# ---------- detect OS / arch ----------
	OS="$(uname -s)"
	ARCH_RAW="$(uname -m)"

	case "$ARCH_RAW" in
	x86_64\|amd64) ARCH="x64" ;;
	aarch64\|arm64) ARCH="arm64" ;;
	s390x) ARCH="s390x" ;;
	*) error "Unsupported architecture: $ARCH_RAW" ;;
	esac

	IS_WINDOWS=0
	IS_MAC=0
	IS_LINUX=0

	case "$OS" in
	Darwin) IS_MAC=1 ;;
	Linux) IS_LINUX=1 ;;
	MINGW\|MSYS\|CYGWIN*) IS_WINDOWS=1 ;;
	*) error "Unsupported OS: $OS" ;;
	esac

	# openEuler is Linux, detect via /etc/os-release
	IS_OPENEULER=0
	if [ "$IS_LINUX" = "1" ] && [ -r /etc/os-release ]; then
	if grep -qi '^ID=.*openeuler' /etc/os-release; then
	IS_OPENEULER=1
	fi
	fi

	# ---------- choose prefix ----------
	if [ -n "${LLAMA_PREFIX:-}" ]; then
	PREFIX="$LLAMA_PREFIX"
	else
	# If /usr/local is writable OR we have sudo, default to /usr/local, else ~/.local
	if [ -w /usr/local ] \|\| [ -n "$SUDO" ]; then
	PREFIX="/usr/local"
	else
	PREFIX="${HOME}/.local"
	fi
	fi

	BIN_DIR="${PREFIX}/bin"
	ROOT_DIR="${PREFIX}/lib/llama.cpp" # keep versioned installs here
	mkdir -p "$TMP_DIR"

	# ---------- fetch release tag ----------
	fetch_release_json() {
	local url="$1"
	local out="$2"
	curl -fsSL -H "Accept: application/vnd.github+json" -H "User-Agent: ${UA}" "$url" > "$out" \
	\|\| error "Failed to fetch release metadata (GitHub API rate limit? try later)"
	}

	RELEASE_JSON="$TMP_DIR/release.json"
	if [ -n "$TAG" ]; then
	status "Fetching release info for tag: $TAG"
	fetch_release_json "${API_BASE}/tags/${TAG}" "$RELEASE_JSON"
	else
	status "Fetching latest release info..."
	fetch_release_json "${API_BASE}/latest" "$RELEASE_JSON"
	TAG="$(grep -m1 '"tag_name"' "$RELEASE_JSON" \| cut -d '"' -f4 \|\| true)"
	[ -n "$TAG" ] \|\| error "Could not parse tag_name from GitHub API response"
	fi

	status "Selected release: $TAG"

	# ---------- auto backend detection ----------
	# Important: Linux prebuilt list here has: cpu/vulkan/rocm7.2 only (no CUDA for Linux in your list)
	# Windows has cuda12/cuda13 variants + cudart
	auto_backend() {
	if [ "$IS_MAC" = "1" ]; then
	# No separate metal variant in your list; prebuilt is "macos-ARCH"
	echo "cpu"
	return 0
	fi

	if [ "$IS_WINDOWS" = "1" ]; then
	# Prefer CUDA if nvidia-smi exists, else vulkan if vulkaninfo exists, else cpu
	if available nvidia-smi; then
	# Pick cuda12 by default (more common driver availability); user can override
	echo "cuda12"
	elif available vulkaninfo; then
	echo "vulkan"
	else
	echo "cpu"
	fi
	return 0
	fi

	# Linux (including openEuler): detect ROCm, then Vulkan, else CPU
	if available rocminfo; then
	echo "rocm"
	elif available vulkaninfo; then
	echo "vulkan"
	else
	echo "cpu"
	fi
	}

	if [ "$BACKEND" = "auto" ]; then
	BACKEND="$(auto_backend)"
	status "Auto-selected backend: $BACKEND"
	else
	status "Requested backend: $BACKEND"
	fi

	# ---------- build asset filename ----------
	ASSET=""
	ASSET2="" # extra (cudart) for some Windows CUDA builds
	EXT="tar.gz"

	if [ "$IS_MAC" = "1" ]; then
	case "$BACKEND" in
	cpu\|vulkan\|rocm) : ;; # tolerate; mac uses same tarball
	xcframework)
	# iOS XCFramework
	EXT="zip"
	ASSET="llama-${TAG}-xcframework.zip"
	;;
	*)
	warning "macOS: backend '$BACKEND' ignored; using macos prebuilt"
	;;
	esac
	if [ -z "$ASSET" ]; then
	ASSET="llama-${TAG}-bin-macos-${ARCH}.tar.gz"
	EXT="tar.gz"
	fi

	elif [ "$IS_WINDOWS" = "1" ]; then
	EXT="zip"
	case "$BACKEND" in
	cpu)
	ASSET="llama-${TAG}-bin-win-cpu-${ARCH}.zip"
	;;
	cuda12)
	[ "$ARCH" = "x64" ] \|\| error "Windows CUDA builds are x64 only in your list"
	ASSET="llama-${TAG}-bin-win-cuda-12.4-x64.zip"
	ASSET2="cudart-llama-bin-win-cuda-12.4-x64.zip"
	;;
	cuda13)
	[ "$ARCH" = "x64" ] \|\| error "Windows CUDA builds are x64 only in your list"
	ASSET="llama-${TAG}-bin-win-cuda-13.1-x64.zip"
	# Your list shows only cudart for 12.4; keep 13 without extra unless they add it later
	;;
	vulkan)
	[ "$ARCH" = "x64" ] \|\| error "Windows Vulkan build in your list is x64 only"
	ASSET="llama-${TAG}-bin-win-vulkan-x64.zip"
	;;
	sycl)
	[ "$ARCH" = "x64" ] \|\| error "Windows SYCL build in your list is x64 only"
	ASSET="llama-${TAG}-bin-win-sycl-x64.zip"
	;;
	hip)
	[ "$ARCH" = "x64" ] \|\| error "Windows HIP Radeon build in your list is x64 only"
	ASSET="llama-${TAG}-bin-win-hip-radeon-x64.zip"
	;;
	opencl-adreno)
	[ "$ARCH" = "arm64" ] \|\| error "Windows OpenCL Adreno build in your list is arm64 only"
	ASSET="llama-${TAG}-bin-win-opencl-adreno-arm64.zip"
	;;
	*)
	error "Unsupported Windows backend: $BACKEND"
	;;
	esac

	else
	# Linux / openEuler
	if [ "$IS_OPENEULER" = "1" ]; then
	# openEuler assets are named differently and don't include x64/arm64 tags in the same style
	# Map ARCH -> x86/aarch64 in names
	OE_ARCH=""
	case "$ARCH" in
	x64) OE_ARCH="x86" ;;
	arm64) OE_ARCH="aarch64" ;;
	*) error "openEuler assets in your list support x86 and aarch64 only; got: $ARCH" ;;
	esac

	case "$OE_FAMILY" in
	310p\|910b) : ;;
	*) error "openEuler: LLAMA_OE_FAMILY must be 310p or 910b" ;;
	esac

	if [ "$OE_ACLGRAPH" = "1" ]; then
	ASSET="llama-${TAG}-bin-${OE_FAMILY}-openEuler-${OE_ARCH}-aclgraph.tar.gz"
	else
	ASSET="llama-${TAG}-bin-${OE_FAMILY}-openEuler-${OE_ARCH}.tar.gz"
	fi
	EXT="tar.gz"

	# Note: backend selection is ignored for openEuler because your assets are family-based
	if [ "$BACKEND" != "cpu" ]; then
	status "openEuler: ignoring backend='$BACKEND' (assets are 310p/910b based)"
	fi

	else
	# Ubuntu assets (Linux)
	case "$BACKEND" in
	cpu)
	if [ "$ARCH" = "s390x" ]; then
	ASSET="llama-${TAG}-bin-ubuntu-s390x.tar.gz"
	else
	[ "$ARCH" = "x64" ] \|\| error "Ubuntu CPU assets in your list are x64 and s390x only"
	ASSET="llama-${TAG}-bin-ubuntu-x64.tar.gz"
	fi
	;;
	vulkan)
	[ "$ARCH" = "x64" ] \|\| error "Ubuntu Vulkan asset in your list is x64 only"
	ASSET="llama-${TAG}-bin-ubuntu-vulkan-x64.tar.gz"
	;;
	rocm)
	[ "$ARCH" = "x64" ] \|\| error "Ubuntu ROCm asset in your list is x64 only"
	ASSET="llama-${TAG}-bin-ubuntu-rocm-7.2-x64.tar.gz"
	;;
	*)
	error "Unsupported Linux backend: $BACKEND (use cpu\|vulkan\|rocm\|auto)"
	;;
	esac
	EXT="tar.gz"
	fi
	fi

	[ -n "$ASSET" ] \|\| error "Internal: failed to choose asset"

	# ---------- download ----------
	dl() {
	local filename="$1"
	local dest="$2"
	local url="https://github.com/${REPO}/releases/download/${TAG}/${filename}"
	status "Downloading: $filename"
	curl --fail --show-error --location --progress-bar -H "User-Agent: ${UA}" \
	-o "$dest" "$url" \|\| error "Download failed: $filename"
	}

	ARCHIVE1="$TMP_DIR/asset1.${EXT}"
	dl "$ASSET" "$ARCHIVE1"

	ARCHIVE2=""
	if [ -n "$ASSET2" ]; then
	ARCHIVE2="$TMP_DIR/asset2.zip"
	dl "$ASSET2" "$ARCHIVE2"
	fi

	# ---------- extract ----------
	EXTRACT_DIR="$TMP_DIR/extract"
	mkdir -p "$EXTRACT_DIR"

	if [ "$EXT" = "zip" ]; then
	NEEDS="$(require unzip)"
	[ -z "$NEEDS" ] \|\| error "Missing required tool: unzip"
	unzip -q "$ARCHIVE1" -d "$EXTRACT_DIR"
	else
	NEEDS="$(require tar)"
	[ -z "$NEEDS" ] \|\| error "Missing required tool: tar"
	tar -xzf "$ARCHIVE1" -C "$EXTRACT_DIR"
	fi

	if [ -n "$ARCHIVE2" ]; then
	NEEDS="$(require unzip)"
	[ -z "$NEEDS" ] \|\| error "Missing required tool: unzip"
	unzip -q "$ARCHIVE2" -d "$EXTRACT_DIR"
	fi

	# ---------- install layout ----------
	# We keep a versioned directory under PREFIX/lib/llama.cpp and then link executables into PREFIX/bin
	INSTALL_ID="${TAG}-$(uname -s \| tr '[:upper:]' '[:lower:]')-${ARCH}-${BACKEND}"
	INSTALL_DIR="${ROOT_DIR}/${INSTALL_ID}"

	status "Installing into: $INSTALL_DIR"
	if [ -n "$SUDO" ] && [ "$PREFIX" = "/usr/local" ]; then
	$SUDO mkdir -p "$BIN_DIR" "$ROOT_DIR"
	$SUDO rm -rf "$INSTALL_DIR"
	$SUDO mkdir -p "$INSTALL_DIR"
	$SUDO cp -R "$EXTRACT_DIR"/. "$INSTALL_DIR"/
	else
	mkdir -p "$BIN_DIR" "$ROOT_DIR"
	rm -rf "$INSTALL_DIR"
	mkdir -p "$INSTALL_DIR"
	cp -R "$EXTRACT_DIR"/. "$INSTALL_DIR"/
	fi

	# ---------- link executables ----------
	link_one() {
	local src="$1"
	local name="$2"
	if [ -n "$SUDO" ] && [ "$PREFIX" = "/usr/local" ]; then
	$SUDO ln -sf "$src" "${BIN_DIR}/${name}" 2>/dev/null \|\| $SUDO cp -f "$src" "${BIN_DIR}/${name}"
	else
	ln -sf "$src" "${BIN_DIR}/${name}" 2>/dev/null \|\| cp -f "$src" "${BIN_DIR}/${name}"
	fi
	}

	status "Linking binaries into: $BIN_DIR"

	if [ "$IS_WINDOWS" = "1" ]; then
	# link .exe and .bat
	wins=()
	while IFS= read -r line; do wins+=("$line"); done < <(find "$INSTALL_DIR" -type f \( -iname '.exe' -o -iname '.bat' \) 2>/dev/null \|\| true)
	[ "${#wins[@]}" -gt 0 ] \|\| warning "No .exe/.bat found. Asset layout may have changed"
	for f in "${wins[@]}"; do
	link_one "$f" "$(basename "$f")"
	done
	else
	# link all executable files (best effort)
	nix=()
	while IFS= read -r line; do nix+=("$line"); done < <(find "$INSTALL_DIR" -type f -perm -111 2>/dev/null \|\| true)
	[ "${#nix[@]}" -gt 0 ] \|\| warning "No executable files found. Asset layout may have changed"
	for f in "${nix[@]}"; do
	link_one "$f" "$(basename "$f")"
	done
	# friendly alias
	if [ -f "${BIN_DIR}/llama-cli" ]; then
	link_one "${BIN_DIR}/llama-cli" "llama"
	fi
	fi

	# ---------- optional systemd (Linux only) ----------
	configure_systemd() {
	[ "$IS_LINUX" = "1" ] \|\| return 0
	[ "$ENABLE_SYSTEMD" = "1" ] \|\| return 0
	available systemctl \|\| { warning "systemctl not found; skipping systemd service"; return 0; }

	local server_path="${BIN_DIR}/llama-server"
	[ -x "$server_path" ] \|\| { warning "llama-server not found in ${BIN_DIR}; skipping systemd"; return 0; }

	if [ -z "$SUDO" ]; then
	warning "No sudo available; cannot install system-wide systemd service. (Set LLAMA_PREFIX=~/.local and use user systemd manually.)"
	return 0
	fi

	status "Creating llama.cpp systemd service (llama-cpp.service)..."
	$SUDO tee /etc/systemd/system/llama-cpp.service >/dev/null <<EOF
	[Unit]
	Description=llama.cpp server
	After=network-online.target
	Wants=network-online.target

	[Service]
	Type=simple
	ExecStart=${server_path} --host ${SERVER_HOST} --port ${SERVER_PORT} ${SERVER_ARGS}
	Restart=always
	RestartSec=2
	Environment="PATH=/usr/local/bin:/usr/bin:/bin:${BIN_DIR}"

	[Install]
	WantedBy=multi-user.target
	EOF

	status "Enabling and starting llama-cpp.service..."
	$SUDO systemctl daemon-reload
	$SUDO systemctl enable llama-cpp.service
	$SUDO systemctl restart llama-cpp.service

	status "llama.cpp server should be listening on ${SERVER_HOST}:${SERVER_PORT}"
	}

	configure_systemd

	status "Install complete"
	echo
	echo "Installed: $INSTALL_DIR"
	echo "Binaries: $BIN_DIR"
	echo
	if [ "$IS_WINDOWS" = "1" ]; then
	echo "Try (Git Bash):"
	echo " ${BIN_DIR}/llama-cli.exe --help"
	echo " ${BIN_DIR}/llama-server.exe --help"
	else
	echo "Try:"
	echo " ${BIN_DIR}/llama-cli --help"
	echo " ${BIN_DIR}/llama-server --help"
	echo
	if [ "$ENABLE_SYSTEMD" = "1" ] && [ "$IS_LINUX" = "1" ] && available systemctl; then
	echo "Service:"
	echo " systemctl status llama-cpp.service"
	fi
	fi
	}

	main "$@"
No results found