Bonsai 1-bit LLM — Installer & Runner

README.md

Copy/Download the script and run it like this:

bashchmod +x bonsai-install.sh

./bonsai-install.sh

What you'll get: What would you like to do?

1. Install + run interactive chat (CLI)
2. Install + run as local server with WebUI
3. Run interactive chat (already installed)
4. Run local server with WebUI (already installed)
5. Change model size (current: 8B)
6. Uninstall q) Quit Key features:

Auto-detects your platform (macOS Apple Silicon/Intel, Linux CUDA/CPU) and builds with the right flags Downloads the correct GGUF from Hugging Face with a progress bar Skips re-downloading/rebuilding if already done Option 2 auto-opens your browser to http://localhost:8080 when the server is ready The server exposes a full OpenAI-compatible API at /v1/chat/completions Supports all three model sizes (1.7B / 4B / 8B) Everything installs to ~/bonsai (clean uninstall via option 6)

Note: On first run, building llama.cpp takes 1–3 minutes. The 8B model download is ~1.15 GB. After that, options 3/4 skip straight to running.

bonsai-install.sh

#!/usr/bin/env bash
# =============================================================================
#  Bonsai 1-bit LLM — Installer & Runner
#  Supports: macOS (Apple Silicon + Intel), Linux (CUDA / CPU)
# =============================================================================

set -euo pipefail

# ── Colours ──────────────────────────────────────────────────────────────────
RED='\033[0;31m'; GREEN='\033[0;32m'; YELLOW='\033[1;33m'
CYAN='\033[0;36m'; BOLD='\033[1m'; RESET='\033[0m'

info()    { echo -e "${CYAN}${BOLD}[INFO]${RESET}  $*"; }
success() { echo -e "${GREEN}${BOLD}[OK]${RESET}    $*"; }
warn()    { echo -e "${YELLOW}${BOLD}[WARN]${RESET}  $*"; }
error()   { echo -e "${RED}${BOLD}[ERROR]${RESET} $*" >&2; exit 1; }
step()    { echo -e "\n${BOLD}━━━  $*  ━━━${RESET}"; }

# ── Defaults ─────────────────────────────────────────────────────────────────
INSTALL_DIR="$HOME/bonsai"
REPO_URL="https://github.com/PrismML-Eng/llama.cpp.git"
MODEL_SIZE="${BONSAI_MODEL:-8B}"          # 1.7B | 4B | 8B
SERVER_PORT="${BONSAI_PORT:-8080}"
CTX_SIZE=65536
GPU_LAYERS=99

# ── Banner ────────────────────────────────────────────────────────────────────
echo -e "
${CYAN}${BOLD}
  ██████╗  ██████╗ ███╗   ██╗███████╗ █████╗ ██╗
  ██╔══██╗██╔═══██╗████╗  ██║██╔════╝██╔══██╗██║
  ██████╔╝██║   ██║██╔██╗ ██║███████╗███████║██║
  ██╔══██╗██║   ██║██║╚██╗██║╚════██║██╔══██║██║
  ██████╔╝╚██████╔╝██║ ╚████║███████║██║  ██║██║
  ╚═════╝  ╚═════╝ ╚═╝  ╚═══╝╚══════╝╚═╝  ╚═╝╚═╝
${RESET}${BOLD}  1-bit LLM Installer & Runner${RESET}
  PrismML • Apache 2.0 • ~1 GB for 8B params
"

# ── Menu ──────────────────────────────────────────────────────────────────────
echo -e "${BOLD}What would you like to do?${RESET}"
echo "  1) Install + run interactive chat (CLI)"
echo "  2) Install + run as local server with WebUI"
echo "  3) Run interactive chat (already installed)"
echo "  4) Run local server with WebUI (already installed)"
echo "  5) Change model size (current: ${MODEL_SIZE})"
echo "  6) Uninstall"
echo "  q) Quit"
echo ""
read -rp "Choice [1]: " CHOICE
CHOICE="${CHOICE:-1}"

# ── Model size picker ─────────────────────────────────────────────────────────
pick_model_size() {
    echo -e "\n${BOLD}Choose model size:${RESET}"
    echo "  1) 8B  — 1.15 GB  (best quality, recommended)"
    echo "  2) 4B  — 0.5 GB   (balanced)"
    echo "  3) 1.7B — 0.24 GB (smallest, Raspberry Pi / edge)"
    read -rp "Choice [1]: " SZ
    case "${SZ:-1}" in
        1) MODEL_SIZE="8B"  ;;
        2) MODEL_SIZE="4B"  ;;
        3) MODEL_SIZE="1.7B" ;;
        *) MODEL_SIZE="8B"  ;;
    esac
    success "Model size set to ${MODEL_SIZE}"
}

if [[ "$CHOICE" == "5" ]]; then
    pick_model_size
    # Re-show menu after picking
    exec "$0"
fi

if [[ "$CHOICE" == "6" ]]; then
    step "Uninstall"
    if [[ -d "$INSTALL_DIR" ]]; then
        read -rp "Remove $INSTALL_DIR? [y/N] " CONFIRM
        [[ "${CONFIRM,,}" == "y" ]] && rm -rf "$INSTALL_DIR" && success "Removed $INSTALL_DIR" || info "Aborted."
    else
        warn "Nothing to remove at $INSTALL_DIR"
    fi
    exit 0
fi

if [[ "$CHOICE" == "q" ]]; then exit 0; fi

# ── Detect platform ───────────────────────────────────────────────────────────
detect_platform() {
    OS="$(uname -s)"
    ARCH="$(uname -m)"

    case "$OS" in
        Darwin)
            PLATFORM="macos"
            if [[ "$ARCH" == "arm64" ]]; then
                BUILD_FLAGS=""          # Metal on by default for Apple Silicon
                HW_DESC="macOS Apple Silicon (Metal)"
            else
                BUILD_FLAGS="-DGGML_METAL=OFF"
                HW_DESC="macOS Intel (CPU)"
            fi
            ;;
        Linux)
            PLATFORM="linux"
            if command -v nvcc &>/dev/null || [[ -d /usr/local/cuda ]]; then
                BUILD_FLAGS="-DGGML_CUDA=ON"
                HW_DESC="Linux + NVIDIA CUDA"
            else
                BUILD_FLAGS=""
                HW_DESC="Linux CPU"
                warn "No CUDA detected — building CPU-only (slower). Install CUDA toolkit for GPU acceleration."
            fi
            ;;
        *)
            error "Unsupported OS: $OS. Use macOS or Linux (for Windows, use WSL2)."
            ;;
    esac

    info "Platform: ${HW_DESC}"
}

# ── Dependency checks ─────────────────────────────────────────────────────────
check_deps() {
    step "Checking dependencies"
    local missing=()

    for cmd in git cmake make; do
        if ! command -v "$cmd" &>/dev/null; then
            missing+=("$cmd")
        fi
    done

    if [[ "$PLATFORM" == "macos" ]] && ! xcode-select -p &>/dev/null; then
        warn "Xcode Command Line Tools not found. Installing..."
        xcode-select --install
        read -rp "Press Enter once Xcode CLT installation completes..."
    fi

    if [[ ${#missing[@]} -gt 0 ]]; then
        error "Missing required tools: ${missing[*]}\n\n  macOS:  brew install ${missing[*]}\n  Ubuntu: sudo apt install ${missing[*]}"
    fi

    success "All dependencies found"
}

# ── HuggingFace model URLs ────────────────────────────────────────────────────
get_model_url() {
    case "$MODEL_SIZE" in
        8B)   echo "https://huggingface.co/prism-ml/Bonsai-8B-gguf/resolve/main/Bonsai-8B-Q1_0.gguf" ;;
        4B)   echo "https://huggingface.co/prism-ml/Bonsai-4B-gguf/resolve/main/Bonsai-4B-Q1_0.gguf" ;;
        1.7B) echo "https://huggingface.co/prism-ml/Bonsai-1.7B-gguf/resolve/main/Bonsai-1.7B-Q1_0.gguf" ;;
        *)    error "Unknown model size: $MODEL_SIZE" ;;
    esac
}

get_model_filename() {
    case "$MODEL_SIZE" in
        8B)   echo "Bonsai-8B-Q1_0.gguf" ;;
        4B)   echo "Bonsai-4B-Q1_0.gguf" ;;
        1.7B) echo "Bonsai-1.7B-Q1_0.gguf" ;;
    esac
}

# ── Install ───────────────────────────────────────────────────────────────────
install_bonsai() {
    step "Installing Bonsai to $INSTALL_DIR"
    mkdir -p "$INSTALL_DIR/models"

    # ── Clone PrismML llama.cpp fork ──────────────────────────────────────────
    if [[ -d "$INSTALL_DIR/llama.cpp/.git" ]]; then
        info "llama.cpp fork already cloned — pulling latest..."
        git -C "$INSTALL_DIR/llama.cpp" pull --ff-only || warn "Could not pull updates (offline?)"
    else
        info "Cloning PrismML llama.cpp fork (includes Q1_0 1-bit kernels)..."
        git clone --depth=1 "$REPO_URL" "$INSTALL_DIR/llama.cpp"
    fi

    # ── Build ──────────────────────────────────────────────────────────────────
    BINARY="$INSTALL_DIR/llama.cpp/build/bin/llama-cli"
    if [[ -f "$BINARY" ]]; then
        info "Binaries already built — skipping build step"
    else
        info "Building llama.cpp (this takes 1–3 minutes)..."
        cmake -S "$INSTALL_DIR/llama.cpp" \
              -B "$INSTALL_DIR/llama.cpp/build" \
              -DCMAKE_BUILD_TYPE=Release \
              $BUILD_FLAGS
        cmake --build "$INSTALL_DIR/llama.cpp/build" \
              --config Release \
              -j "$(getconf _NPROCESSORS_ONLN 2>/dev/null || echo 4)"
        success "Build complete"
    fi

    # ── Download model ─────────────────────────────────────────────────────────
    MODEL_FILE="$(get_model_filename)"
    MODEL_PATH="$INSTALL_DIR/models/$MODEL_FILE"

    if [[ -f "$MODEL_PATH" ]]; then
        success "Model already downloaded: $MODEL_FILE"
    else
        MODEL_URL="$(get_model_url)"
        info "Downloading $MODEL_FILE from Hugging Face..."
        info "URL: $MODEL_URL"
        echo ""

        if command -v wget &>/dev/null; then
            wget --show-progress -O "$MODEL_PATH" "$MODEL_URL"
        elif command -v curl &>/dev/null; then
            curl -L --progress-bar -o "$MODEL_PATH" "$MODEL_URL"
        else
            error "Neither wget nor curl found. Please install one and retry."
        fi

        success "Model downloaded: $MODEL_PATH"
    fi
}

# ── Run: interactive CLI ──────────────────────────────────────────────────────
run_chat() {
    MODEL_PATH="$INSTALL_DIR/models/$(get_model_filename)"
    BINARY="$INSTALL_DIR/llama.cpp/build/bin/llama-cli"

    [[ -f "$BINARY" ]]     || error "Binary not found. Run install first (option 1 or 2)."
    [[ -f "$MODEL_PATH" ]] || error "Model not found at $MODEL_PATH. Run install first."

    step "Starting Bonsai ${MODEL_SIZE} — Interactive Chat"
    echo -e "  ${YELLOW}Type your message and press Enter. Ctrl+C to exit.${RESET}\n"

    "$BINARY" \
        -m "$MODEL_PATH" \
        --interactive-first \
        --color \
        -n -1 \
        --temp 0.5 \
        --top-p 0.85 \
        --top-k 20 \
        --repeat-penalty 1.1 \
        -ngl "$GPU_LAYERS" \
        --ctx-size 4096 \
        -p "You are a helpful assistant."
}

# ── Run: local server + WebUI ─────────────────────────────────────────────────
run_server() {
    MODEL_PATH="$INSTALL_DIR/models/$(get_model_filename)"
    BINARY="$INSTALL_DIR/llama.cpp/build/bin/llama-server"

    [[ -f "$BINARY" ]]     || error "Server binary not found. Run install first (option 1 or 2)."
    [[ -f "$MODEL_PATH" ]] || error "Model not found at $MODEL_PATH. Run install first."

    step "Starting Bonsai ${MODEL_SIZE} — Local Server"
    echo ""
    echo -e "  ${GREEN}${BOLD}WebUI:${RESET}   http://localhost:${SERVER_PORT}"
    echo -e "  ${GREEN}${BOLD}API:${RESET}     http://localhost:${SERVER_PORT}/v1/chat/completions"
    echo -e "  ${YELLOW}Ctrl+C to stop the server${RESET}"
    echo ""

    # Open browser after a short delay (best-effort)
    (
        sleep 3
        if command -v open &>/dev/null; then
            open "http://localhost:${SERVER_PORT}"
        elif command -v xdg-open &>/dev/null; then
            xdg-open "http://localhost:${SERVER_PORT}" &>/dev/null
        fi
    ) &

    "$BINARY" \
        -m "$MODEL_PATH" \
        --host 0.0.0.0 \
        --port "$SERVER_PORT" \
        -ngl "$GPU_LAYERS" \
        --ctx-size "$CTX_SIZE" \
        --temp 0.5 \
        --top-p 0.85 \
        --top-k 20
}

# ── Main dispatcher ───────────────────────────────────────────────────────────
detect_platform
check_deps

case "$CHOICE" in
    1)
        install_bonsai
        echo ""
        run_chat
        ;;
    2)
        install_bonsai
        echo ""
        run_server
        ;;
    3)
        run_chat
        ;;
    4)
        run_server
        ;;
    *)
        error "Invalid choice: $CHOICE"
        ;;
esac