Upload installpi.sh

installpi.sh

@@ -0,0 +1,642 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# ==============================================================================
+# installpi.sh - 一鍵全自動安裝 pi + llama.cpp + Qwen3.6 35B A3B MTP 模型
+# 特性：無 sudo、無互動、全自動、可斷點續傳、安裝後自動驗證
+# 模型：unsloth/Qwen3.6-35B-A3B-MTP-GGUF (優先 UD-Q4_K_M, 啟用 draft-mtp 加速)
+# 參考: https://huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF
+# ==============================================================================
+
+HOME_DIR="$HOME"
+LLAMA_DIR="$HOME_DIR/llama-cpp"
+LLAMA_BIN="$LLAMA_DIR/bin"
+MODEL_DIR="$HOME_DIR/models"
+PI_MODELS_JSON="$HOME_DIR/.pi/agent/models.json"
+LLAMA_PORT=8081
+REPO="unsloth/Qwen3.6-35B-A3B-MTP-GGUF"
+MODEL_ID="qwen3.6-35b-a3b"
+PROVIDER_ID="llama-cpp-local"
+
+# 基礎輸出函數
+info()  { echo -e "\033[36m[INFO]\033[0m $*"; }
+ok()    { echo -e "\033[32m[OK]\033[0m $*"; }
+warn()  { echo -e "\033[33m[WARN]\033[0m $*" >&2; }
+error() { echo -e "\033[31m[ERROR]\033[0m $*" >&2; exit 1; }
+
+# 確保必要目錄
+mkdir -p "$LLAMA_BIN" "$MODEL_DIR" "$(dirname "$PI_MODELS_JSON")" "$HOME_DIR/.local/bin"
+
+# 檢查必要工具
+for cmd in curl python3 tar; do
+    command -v "$cmd" >/dev/null 2>&1 || error "缺少必要工具: $cmd，請先安裝"
+done
+
+# ==============================================================================
+# 0. 確保 Node.js >=22.19.0 (pi 依賴，避免 pi 安裝腳本進入互動模式)
+# ==============================================================================
+ensure_node() {
+    if command -v node >/dev/null 2>&1; then
+        if node -e 'const [maj,min,patch]=process.versions.node.split(".").map(Number); process.exit(maj>22||(maj===22&&(min>19||(min===19&&patch>=0)))?0:1)' >/dev/null 2>&1; then
+            ok "Node.js 已滿足: $(node --version)"
+            return 0
+        else
+            warn "Node.js 版本過舊: $(node --version)，將自動安裝新版..."
+        fi
+    else
+        info "未檢測到 Node.js，將自動安裝..."
+    fi
+
+    # 採用 standalone 方式安裝 Node 22.x 到 ~/.local/share/pi-node (無 sudo)
+    info "自動安裝 Node.js 22.x (standalone 模式)..."
+    NODE_DIST_BASE="https://nodejs.org/dist/latest-v22.x"
+    NODE_BASE_DIR="${XDG_DATA_HOME:-$HOME/.local/share}/pi-node"
+    NODE_TMP_DIR="/tmp/pi-node-$$"
+    mkdir -p "$NODE_TMP_DIR" "$NODE_BASE_DIR"
+
+    # 判斷平台 / 架構
+    OS="$(uname -s)"
+    ARCH="$(uname -m)"
+    case "$OS" in Linux) PLATFORM="linux" ;; Darwin) PLATFORM="darwin" ;; *) error "不支援的 OS: $OS" ;; esac
+    case "$ARCH" in x86_64|amd64) NODE_ARCH="x64" ;; aarch64|arm64) NODE_ARCH="arm64" ;; *) error "不支援的架構: $ARCH" ;; esac
+
+    info "解析 Node.js 下載連結 ($PLATFORM-$NODE_ARCH)..."
+    curl -fsSL --retry 3 --retry-delay 5 "$NODE_DIST_BASE/SHASUMS256.txt" -o "$NODE_TMP_DIR/SHASUMS256.txt" || error "下載 SHASUMS 失敗"
+
+    NODE_FILE=$(awk -v suffix="-$PLATFORM-$NODE_ARCH.tar.xz" 'index($2,"node-v")==1 && substr($2,length($2)-length(suffix)+1)==suffix{print $2; exit}' "$NODE_TMP_DIR/SHASUMS256.txt")
+    [[ -n "$NODE_FILE" ]] || error "找不到對應平台的 Node.js 二進位"
+
+    info "下載 Node.js $NODE_FILE ..."
+    curl -fsSL --retry 3 --retry-delay 5 "$NODE_DIST_BASE/$NODE_FILE" -o "$NODE_TMP_DIR/$NODE_FILE" || error "Node.js 下載失敗"
+
+    # 校驗
+    if command -v sha256sum >/dev/null 2>&1; then
+        (cd "$NODE_TMP_DIR" && grep -F "$NODE_FILE" SHASUMS256.txt > SHASUMS256.selected && sha256sum -c SHASUMS256.selected) || warn "SHA256 校驗失敗，繼續安裝..."
+    fi
+
+    # 需要 xz 解壓
+    if ! command -v xz >/dev/null 2>&1; then
+        warn "未找到 xz，將嘗試使用 tar 自動解壓 (若失敗請手動安裝 xz-utils)"
+    fi
+
+    NODE_EXTRACT_DIR="$NODE_BASE_DIR/${NODE_FILE%.tar.xz}"
+    rm -rf "$NODE_EXTRACT_DIR"
+    tar -xf "$NODE_TMP_DIR/$NODE_FILE" -C "$NODE_BASE_DIR" || error "解壓 Node.js 失敗"
+    rm -f "$NODE_BASE_DIR/current"
+    ln -s "$NODE_EXTRACT_DIR" "$NODE_BASE_DIR/current"
+    rm -rf "$NODE_TMP_DIR"
+
+    export PATH="$NODE_BASE_DIR/current/bin:$PATH"
+    # 寫入 bashrc / zshrc
+    for rc in "$HOME/.bashrc" "$HOME/.zshrc" "$HOME/.profile"; do
+        if [[ -f "$rc" ]] || [[ "$rc" == "$HOME/.bashrc" ]]; then
+            grep -q "pi-node/current/bin" "$rc" 2>/dev/null || echo "export PATH=\"$NODE_BASE_DIR/current/bin:\$PATH\"" >> "$rc"
+        fi
+    done
+
+    if command -v node >/dev/null 2>&1; then
+        ok "Node.js 安裝完成: $(node --version) at $(command -v node)"
+    else
+        error "Node.js 安裝後仍無法找到，請檢查 PATH"
+    fi
+}
+ensure_node
+
+# 確保 npm 可用
+command -v npm >/dev/null 2>&1 || error "npm 仍不可用，Node 安裝可能失敗"
+info "npm 版本: $(npm --version)"
+
+
+# ==============================================================================
+# 1. 下載 llama.cpp 預編譯二進位
+# ==============================================================================
+info "1) 下載 llama.cpp 預編譯二進位檔..."
+LLAMA_TAG="b10545"
+LLAMA_URL="https://github.com/ggml-org/llama.cpp/releases/download/${LLAMA_TAG}/llama-${LLAMA_TAG}-bin-ubuntu-x64.tar.gz"
+if [[ ! -x "$LLAMA_BIN/llama-server" ]]; then
+    info "下載並解壓 llama.cpp ${LLAMA_TAG}..."
+    curl -fL --retry 3 --retry-delay 5 --progress-bar "$LLAMA_URL" -o /tmp/llama.tar.gz || error "llama.cpp 下載失敗"
+    tar -xzf /tmp/llama.tar.gz -C /tmp || error "解壓失敗"
+    SRC_DIR=$(find /tmp -maxdepth 1 -type d -name "llama-b*" | head -n 1)
+    if [[ -n "$SRC_DIR" ]]; then
+        cp -a "$SRC_DIR"/llama-* "$LLAMA_BIN/" 2>/dev/null || true
+        cp -a "$SRC_DIR"/*.so* "$LLAMA_BIN/" 2>/dev/null || true
+        chmod +x "$LLAMA_BIN"/llama-server "$LLAMA_BIN"/llama-cli 2>/dev/null || true
+        rm -rf /tmp/llama.tar.gz "$SRC_DIR"
+        [[ -x "$LLAMA_BIN/llama-server" ]] || error "llama-server 安裝失敗"
+        ok "llama.cpp 安裝完成"
+    else
+        error "解壓 llama.cpp 失敗，找不到解壓目錄"
+    fi
+else
+    ok "llama.cpp 已存在，跳過下載"
+fi
+
+# 更新 PATH
+export PATH="$LLAMA_BIN:$HOME_DIR/.local/bin:$PATH"
+export LD_LIBRARY_PATH="$LLAMA_BIN:${LD_LIBRARY_PATH:-}"
+# 冪等寫入 rc
+for rc in "$HOME/.bashrc" "$HOME/.zshrc"; do
+    touch "$rc" 2>/dev/null || true
+    grep -q "llama-cpp/bin" "$rc" 2>/dev/null || echo "export PATH=\"$LLAMA_BIN:\$PATH\"" >> "$rc"
+    grep -q "LD_LIBRARY_PATH.*llama-cpp" "$rc" 2>/dev/null || echo "export LD_LIBRARY_PATH=\"$LLAMA_BIN:\$LD_LIBRARY_PATH\"" >> "$rc"
+done
+
+# ==============================================================================
+# 2. 使用 curl 下載模型 (unsloth/Qwen3.6-35B-A3B-MTP-GGUF)
+# ==============================================================================
+info "2) 使用 curl 下載模型 ${REPO} (支援斷點續傳)..."
+
+# 使用 Python 獲取 gguf 清單，明確優先順序：針對 Qwen3.6-35B-A3B-MTP 優先 UD-Q4_K_M / UD-Q4_K_XL
+info "正在獲取模型檔案列表..."
+FILE_LIST=$(python3 -c "
+import urllib.request, json, sys
+repo = '${REPO}'
+url = f'https://huggingface.co/api/models/{repo}/tree/main'
+try:
+    with urllib.request.urlopen(url, timeout=30) as resp:
+        data = json.load(resp)
+    files = [item['path'] for item in data if item.get('type')=='file' and item['path'].endswith('.gguf')]
+    # 排除 mmproj (視覺投影) - Qwen3.6 文本模型不需要，若需要可另外處理
+    # 優先順序：UD-Q4_K_M 為 app.py 預設，其次 UD-Q4_K_XL，再其他
+    def score(f):
+        fl = f.lower()
+        if 'ud-q4_k_m' in fl: return 0
+        if 'ud-q4_k_xl' in fl: return 1
+        if 'q4_k_m' in fl: return 2
+        if 'q4_k_xl' in fl: return 3
+        if 'q4' in fl: return 4
+        if 'q5' in fl or 'q6' in fl or 'q8' in fl: return 5
+        if 'q3' in fl or 'iq' in fl or 'q2' in fl: return 6
+        return 10
+    # 對於 MTP 倉庫，主模型本身已含 MTP heads，無需單獨 mtp 檔案；僅排除 mmproj
+    # 早期 gemma 倉庫有獨立 mtp 檔，保留相容邏輯但對 Qwen3.6 MTP 無獨立檔
+    main_candidates = [f for f in files if 'mmproj' not in f.lower()]
+    # 對 MTP 倉庫：優先選含 MTP 的同名檔案其實就是主模型；若倉庫名含 MTP，全部主模型皆為 MTP 版本
+    is_mtp_repo = 'mtp' in repo.lower()
+    if not is_mtp_repo:
+        main_candidates = [f for f in main_candidates if 'mtp' not in f.lower()]
+    main_candidates.sort(key=score)
+    main = main_candidates[0] if main_candidates else ''
+    # mtp 單獨檔 (僅非 MTP 倉庫需要)
+    mtp = ''
+    if not is_mtp_repo:
+        mtp = next((f for f in files if 'mtp' in f.lower()), '')
+    # 若是 MTP 倉庫，標記主模型即為 MTP 版本
+    mtp_marker = 'MTP integrated' if is_mtp_repo else mtp
+    print(f'MAIN:{main}')
+    print(f'MTP:{mtp}')
+    print(f'MTP_MARKER:{mtp_marker}')
+    print(f'ALL:{ \",\".join(files)}', file=sys.stderr)
+except Exception as e:
+    print(f'ERROR:{e}')
+    sys.exit(1)
+" 2>/tmp/hf_file_list.log)
+
+echo "$FILE_LIST"
+cat /tmp/hf_file_list.log >&2 || true
+
+MAIN_FILE=$(echo "$FILE_LIST" | grep "^MAIN:" | cut -d: -f2- | xargs)
+MTP_FILE=$(echo "$FILE_LIST" | grep "^MTP:" | cut -d: -f2- | xargs)
+
+if [[ -z "$MAIN_FILE" ]]; then
+    error "無法獲取主模型檔案，請檢查網路或 Repo: $REPO (詳見 /tmp/hf_file_list.log)"
+fi
+
+info "找到主模型: $MAIN_FILE"
+DOWNLOAD_URL="https://huggingface.co/${REPO}/resolve/main/${MAIN_FILE}"
+
+if [[ ! -f "$MODEL_DIR/$MAIN_FILE" ]]; then
+    info "開始下載主模型 (支援斷點續傳，若中斷可重新執行此腳本)..."
+    # -C - 斷點續傳, -L 跟隨重定向, --retry 自動重試
+    curl -L -C - --retry 3 --retry-delay 5 --progress-bar "$DOWNLOAD_URL" -o "$MODEL_DIR/$MAIN_FILE" || error "主模型下載失敗"
+    ok "主模型下載完成: $MODEL_DIR/$MAIN_FILE"
+else
+    # 檢查檔案大小是否合理 (>1GB)
+    SIZE=$(stat -c%s "$MODEL_DIR/$MAIN_FILE" 2>/dev/null || stat -f%z "$MODEL_DIR/$MAIN_FILE" 2>/dev/null || echo 0)
+    if [[ "$SIZE" -lt 1000000000 ]]; then
+        warn "主模型檔案過小 (${SIZE} bytes)，可能下載不完整，將嘗試續傳..."
+        curl -L -C - --retry 3 --retry-delay 5 --progress-bar "$DOWNLOAD_URL" -o "$MODEL_DIR/$MAIN_FILE" || error "續傳失敗"
+    fi
+    ok "主模型已存在，跳過下載: $MODEL_DIR/$MAIN_FILE ($SIZE bytes)"
+fi
+MODEL_PATH="$MODEL_DIR/$MAIN_FILE"
+
+MTP_PATH=""
+if [[ -n "$MTP_FILE" ]]; then
+    info "找到 MTP 加速模型: $MTP_FILE"
+    MTP_DOWNLOAD_URL="https://huggingface.co/${REPO}/resolve/main/${MTP_FILE}"
+    if [[ ! -f "$MODEL_DIR/$MTP_FILE" ]]; then
+        info "開始下載 MTP 模型..."
+        curl -L -C - --retry 3 --retry-delay 5 --progress-bar "$MTP_DOWNLOAD_URL" -o "$MODEL_DIR/$MTP_FILE" || warn "MTP 模型下載失敗，將僅使用主模型"
+    else
+        ok "MTP 模型已存在，跳過下載"
+    fi
+    if [[ -f "$MODEL_DIR/$MTP_FILE" ]]; then
+        MTP_PATH="$MODEL_DIR/$MTP_FILE"
+    fi
+else
+    info "未找到 MTP 模型，將僅使用主模型"
+fi
+
+# ==============================================================================
+# 3. 啟動 llama-server
+# ==============================================================================
+info "3) 啟動 llama-server (port $LLAMA_PORT)..."
+# 清理舊進程
+if [[ -f "$HOME_DIR/llama-server.pid" ]]; then
+    OLD_PID=$(cat "$HOME_DIR/llama-server.pid" 2>/dev/null || echo "")
+    if [[ -n "$OLD_PID" ]] && kill -0 "$OLD_PID" 2>/dev/null; then
+        info "停止舊 llama-server (PID $OLD_PID)..."
+        kill "$OLD_PID" 2>/dev/null || true
+        sleep 2
+        kill -9 "$OLD_PID" 2>/dev/null || true
+    fi
+fi
+pkill -f "llama-server.*${LLAMA_PORT}" 2>/dev/null || true
+sleep 1
+
+THREADS=$(nproc 2>/dev/null || echo 8)
+[[ "$THREADS" -gt 16 ]] && THREADS=16
+
+CMD=("$LLAMA_BIN/llama-server"
+    -m "$MODEL_PATH"
+    --host 0.0.0.0
+    --port "$LLAMA_PORT"
+    -c 8192
+    -t "$THREADS"
+    --flash-attn on
+)
+
+# MTP 加速：若為 MTP 倉庫 (unsloth/*-MTP-GGUF) 且 llama-server 支援 spec-type，則啟用 draft-mtp
+# Qwen3.6-35B-A3B-MTP 官方建議 --spec-draft-n-max 2 , Qwen3.5-4B-MTP 為 6
+IS_MTP_REPO=0
+if [[ "$REPO" == *"-MTP-"* ]] || [[ "$REPO" == *"-MTP"* ]]; then
+    IS_MTP_REPO=1
+fi
+if [[ "$IS_MTP_REPO" -eq 1 ]] && "$LLAMA_BIN/llama-server" --help 2>&1 | grep -q "spec-type"; then
+    # 根據模型大小選擇 n-max：35B 用 2，4B 用 6，向後兼容
+    SPEC_N_MAX=2
+    if [[ "$MODEL_PATH" == *"4B"* ]]; then SPEC_N_MAX=6; fi
+    if [[ -n "$MTP_PATH" ]]; then
+        # 舊 gemma 有獨立 mtp 檔，需 --model-draft
+        CMD+=(--spec-type draft-mtp --spec-draft-n-max "$SPEC_N_MAX" --model-draft "$MTP_PATH")
+    else
+        # Qwen3.6 MTP 為單檔整合 MTP heads，直接啟用
+        CMD+=(--spec-type draft-mtp --spec-draft-n-max "$SPEC_N_MAX")
+    fi
+    info "✅ 已啟用 MTP speculative decoding 加速 (--spec-type draft-mtp --spec-draft-n-max $SPEC_N_MAX)"
+elif [[ -n "$MTP_PATH" ]]; then
+    info "MTP 模型已下載至 $MTP_PATH (預設未啟用，如需加速可手動啟用 --spec-type draft-mtp)"
+fi
+
+info "啟動命令: ${CMD[*]}"
+# 確保 log 可寫
+touch "$HOME_DIR/llama-server.log" 2>/dev/null || true
+nohup "${CMD[@]}" > "$HOME_DIR/llama-server.log" 2>&1 &
+echo $! > "$HOME_DIR/llama-server.pid"
+info "llama-server PID: $(cat "$HOME_DIR/llama-server.pid")"
+
+# 等待伺服器就緒 (最長 5 分鐘)
+info "等待 llama-server 就緒 (最多 5 分鐘)..."
+READY=0
+for i in {1..60}; do
+    if curl -sf "http://127.0.0.1:${LLAMA_PORT}/health" >/dev/null 2>&1; then
+        ok "llama-server 已就緒 (http://127.0.0.1:${LLAMA_PORT})"
+        READY=1
+        break
+    fi
+    # 也嘗試 /v1/models 作為備用健康檢查
+    if curl -sf "http://127.0.0.1:${LLAMA_PORT}/v1/models" >/dev/null 2>&1; then
+        ok "llama-server 已就緒 (via /v1/models)"
+        READY=1
+        break
+    fi
+    if ! kill -0 $(cat "$HOME_DIR/llama-server.pid" 2>/dev/null || echo 0) 2>/dev/null; then
+        error "llama-server 啟動失敗，請查看 ~/llama-server.log (最後 50 行):\n$(tail -n 50 "$HOME_DIR/llama-server.log" 2>/dev/null)"
+    fi
+    echo -n "."
+    sleep 5
+done
+echo ""
+if [[ "$READY" -ne 1 ]]; then
+    error "llama-server 在 5 分鐘內未就緒，請查看 ~/llama-server.log"
+fi
+
+# 額外確認 OpenAI 相容端點
+if curl -sf "http://127.0.0.1:${LLAMA_PORT}/v1/models" >/dev/null 2>&1; then
+    ok "OpenAI API 端點正常"
+else
+    warn "OpenAI API 端點未回應，但 health 檢查已通過"
+fi
+
+# ==============================================================================
+# 4. 安裝 pi agent (全自動，無互動)
+# ==============================================================================
+info "4) 安裝 pi agent (全自動)..."
+
+# 確保 npm prefix 可寫，避免需要 sudo
+NPM_PREFIX=$(npm prefix -g 2>/dev/null || npm config get prefix 2>/dev/null || echo "$HOME/.local")
+info "npm 全域前綴: $NPM_PREFIX"
+
+# 如果 pi 已安裝，先檢查版本
+NEED_INSTALL=1
+if command -v pi >/dev/null 2>&1; then
+    ok "pi 已安裝: $(pi --version 2>&1 | head -n 1)"
+    # 嘗試更新到最新 (非強制)
+    NEED_INSTALL=0
+    # 若明確需要重裝，可設定 PI_FORCE_REINSTALL=1
+    if [[ "${PI_FORCE_REINSTALL:-0}" == "1" ]]; then
+        info "PI_FORCE_REINSTALL=1，強制重新安裝..."
+        NEED_INSTALL=1
+    fi
+fi
+
+if [[ "$NEED_INSTALL" -eq 1 ]]; then
+    info "執行 pi 官方安裝腳本 (全自動模式)..."
+
+    # 關鍵：確保無 TTY 互動，pi 安裝腳本在無 TTY 時會自動選擇 install/reinstall
+    # 同時透過環境變數避免任何提示
+    # 若 npm prefix 不可寫，腳本會自動使用 ~/.local
+
+    # 方法1：直接透過 npm 安裝 (最可靠、全自動、無互動)
+    # 使用與官方腳本相同的參數
+    PI_PACKAGE="@earendil-works/pi-coding-agent"
+    # 偵測安裝前綴
+    if npm prefix -g 2>/dev/null | xargs test -w 2>/dev/null; then
+        INSTALL_PREFIX=""
+    else
+        INSTALL_PREFIX="$HOME/.local"
+        info "全域 prefix 不可寫，將安裝到 $INSTALL_PREFIX"
+    fi
+
+    set +e
+    if [[ -n "$INSTALL_PREFIX" ]]; then
+        npm install -g --ignore-scripts --no-fund --no-audit --progress=false --prefix "$INSTALL_PREFIX" "$PI_PACKAGE" 2>&1 | tee /tmp/pi-install.log
+        NPM_STATUS=${PIPESTATUS[0]}
+    else
+        npm install -g --ignore-scripts --no-fund --no-audit --progress=false "$PI_PACKAGE" 2>&1 | tee /tmp/pi-install.log
+        NPM_STATUS=${PIPESTATUS[0]}
+    fi
+    set -e
+
+    if [[ "$NPM_STATUS" -ne 0 ]]; then
+        warn "npm 安裝 pi 失敗 (code $NPM_STATUS)，嘗試備用方法: curl 官方腳本..."
+        # 備用：使用官方腳本，並透過無 TTY 自動安裝
+        # 強制無互動：將 stdin 重定向為空，依賴腳本的「無終端時自動安裝」邏輯
+        curl -fsSL https://pi.dev/install.sh | bash -s -- 2>&1 | tee -a /tmp/pi-install.log || true
+    fi
+
+    # 再次檢查
+    # 更新 PATH 以包含可能的安裝位置
+    export PATH="$HOME/.local/bin:$HOME/.pi/bin:$NPM_PREFIX/bin:$PATH"
+    if [[ -n "${INSTALL_PREFIX:-}" ]]; then
+        export PATH="$INSTALL_PREFIX/bin:$PATH"
+    fi
+    # 也包含 standalone node 的情況
+    NODE_BIN_DIR="${XDG_DATA_HOME:-$HOME/.local/share}/pi-node/current/bin"
+    export PATH="$NODE_BIN_DIR:$PATH"
+
+    hash -r 2>/dev/null || true
+fi
+
+# 最終確認 pi 可用
+PI_BIN=$(command -v pi 2>/dev/null || echo "")
+if [[ -z "$PI_BIN" ]]; then
+    # 嘗試常見路徑
+    for cand in "$HOME/.local/bin/pi" "$HOME/.pi/bin/pi" "$NPM_PREFIX/bin/pi" "/usr/local/bin/pi"; do
+        if [[ -x "$cand" ]]; then
+            PI_BIN="$cand"
+            export PATH="$(dirname "$cand"):$PATH"
+            break
+        fi
+    done
+fi
+
+if [[ -z "$PI_BIN" ]] || [[ ! -x "$PI_BIN" ]]; then
+    error "pi 安裝後仍找不到執行檔，請查看 /tmp/pi-install.log"
+fi
+
+ok "pi 可用: $PI_BIN ($(pi --version 2>&1 | head -n 1))"
+
+# 確保 PATH 寫入 rc
+for rc in "$HOME/.bashrc" "$HOME/.zshrc" "$HOME/.profile"; do
+    touch "$rc" 2>/dev/null || true
+    # pi bin 目錄
+    PI_BIN_DIR="$(dirname "$PI_BIN")"
+    if [[ "$PI_BIN_DIR" != "" ]]; then
+        grep -qF "$PI_BIN_DIR" "$rc" 2>/dev/null || echo "export PATH=\"$PI_BIN_DIR:\$PATH\"" >> "$rc"
+    fi
+    # 也確保 ~/.local/bin 在 PATH
+    grep -qF ".local/bin" "$rc" 2>/dev/null || echo "export PATH=\"\$HOME/.local/bin:\$PATH\"" >> "$rc"
+done
+
+# ==============================================================================
+# 5. 配置 pi 使用本地模型
+# ==============================================================================
+info "5) 配置 pi models.json ($PI_MODELS_JSON)..."
+
+# 確保 pi 設定目錄存在
+mkdir -p "$(dirname "$PI_MODELS_JSON")"
+
+python3 - "$PI_MODELS_JSON" "$LLAMA_PORT" "$MODEL_ID" "$PROVIDER_ID" << 'PY'
+import json, os, sys
+path, port, model_id, provider = sys.argv[1], sys.argv[2], sys.argv[3], sys.argv[4]
+os.makedirs(os.path.dirname(path), exist_ok=True)
+cfg = {"providers": {}}
+if os.path.exists(path):
+    try:
+        with open(path) as f:
+            loaded = json.load(f)
+            if isinstance(loaded, dict):
+                cfg = loaded
+                if "providers" not in cfg:
+                    cfg["providers"] = {}
+    except Exception as e:
+        print(f"[WARN] 無法解析舊 models.json，將覆蓋: {e}")
+
+cfg["providers"][provider] = {
+    "baseUrl": f"http://127.0.0.1:{port}/v1",
+    "api": "openai-completions",
+    "apiKey": "not-needed",
+    "compat": {"supportsDeveloperRole": True, "supportsReasoningEffort": False},
+    "models": [{
+        "id": model_id,
+        "name": "Qwen3.6 35B A3B MTP (Local)",
+        "reasoning": True,
+        "input": ["text", "image"],
+        "contextWindow": 16384,
+        "maxTokens": 8192,
+        "cost": {"input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0}
+    }]
+}
+with open(path, "w") as f:
+    json.dump(cfg, f, indent=2)
+print(f"[OK] 已寫入 {path}")
+print(json.dumps(cfg, indent=2))
+PY
+
+ok "pi 配置完成"
+
+# 建立 alias 和 wrapper (冪等)
+WRAPPER="$HOME/.local/bin/pi-local"
+mkdir -p "$(dirname "$WRAPPER")"
+cat > "$WRAPPER" << EOS
+#!/usr/bin/env bash
+exec pi --provider $PROVIDER_ID --model $MODEL_ID "\$@"
+EOS
+chmod +x "$WRAPPER"
+ok "已建立 pi-local wrapper: $WRAPPER"
+
+# 寫入 alias (避免重複)
+for rc in "$HOME/.bashrc" "$HOME/.zshrc"; do
+    touch "$rc" 2>/dev/null || true
+    grep -q "alias pi-local" "$rc" 2>/dev/null || echo "alias pi-local=\"pi --provider $PROVIDER_ID --model $MODEL_ID\"" >> "$rc"
+done
+
+# ==============================================================================
+# 6. 安裝後驗證 (必須通過才算安裝成功)
+# ==============================================================================
+info "6) 安裝後驗證..."
+
+VERIFICATION_FAILED=0
+
+# 6.1 檢查 llama-server
+info "[驗證 1/4] 檢查 llama-server 健康狀態..."
+if curl -sf "http://127.0.0.1:${LLAMA_PORT}/health" >/dev/null 2>&1; then
+    ok "llama-server 健康檢查通過"
+elif curl -sf "http://127.0.0.1:${LLAMA_PORT}/v1/models" >/dev/null 2>&1; then
+    ok "llama-server (via /v1/models) 健康檢查通過"
+else
+    warn "llama-server 健康檢查失敗"
+    VERIFICATION_FAILED=1
+fi
+
+# 6.2 檢查 OpenAI API 實際推理
+info "[驗證 2/4] 測試 OpenAI 相容 API 推理..."
+API_TEST_RESPONSE=$(curl -sf "http://127.0.0.1:${LLAMA_PORT}/v1/chat/completions" \
+    -H "Content-Type: application/json" \
+    -d '{
+        "model": "'$MODEL_ID'",
+        "messages": [{"role": "user", "content": "用一句話介紹你自己，不要超過20字"}],
+        "max_tokens": 50,
+        "temperature": 0.7
+    }' 2>&1 || echo "CURL_FAILED")
+
+if echo "$API_TEST_RESPONSE" | grep -q '"choices"' && echo "$API_TEST_RESPONSE" | grep -q '"content"'; then
+    ok "OpenAI API 推理測試通過"
+    echo "$API_TEST_RESPONSE" | python3 -m json.tool 2>/dev/null | head -n 30 || echo "$API_TEST_RESPONSE" | head -n 20
+elif echo "$API_TEST_RESPONSE" | grep -q "CURL_FAILED"; then
+    warn "API 推理 curl 失敗: $API_TEST_RESPONSE"
+    VERIFICATION_FAILED=1
+else
+    warn "API 推理回應異常，可能是模型仍在載入中..."
+    echo "$API_TEST_RESPONSE" | head -n 50 >&2
+    # 不立即標記失敗，等待 pi-local 測試
+fi
+
+# 6.3 檢查 pi 是否可用
+info "[驗證 3/4] 檢查 pi 指令..."
+if pi --version >/dev/null 2>&1; then
+    ok "pi --version: $(pi --version 2>&1 | head -n 1)"
+else
+    warn "pi --version 執行失敗"
+    VERIFICATION_FAILED=1
+fi
+
+if pi --help >/dev/null 2>&1; then
+    ok "pi --help 正常"
+fi
+
+# 檢查 models.json 是否正確
+if python3 -c "import json; d=json.load(open('$PI_MODELS_JSON')); assert '$PROVIDER_ID' in d.get('providers',{}); assert any(m['id']=='$MODEL_ID' for m in d['providers']['$PROVIDER_ID'].get('models',[]))" 2>/dev/null; then
+    ok "pi models.json 配置正確"
+else
+    warn "pi models.json 配置檢查失敗"
+    cat "$PI_MODELS_JSON" 2>/dev/null || echo "檔案不存在"
+    VERIFICATION_FAILED=1
+fi
+
+if [[ -x "$WRAPPER" ]]; then
+    ok "pi-local wrapper 存在且可執行"
+else
+    warn "pi-local wrapper 缺失"
+    VERIFICATION_FAILED=1
+fi
+
+# 6.4 測試 pi-local 實際可用 (最關鍵)
+info "[驗證 4/4] 測試 pi-local 實際推理 (超時 90 秒)..."
+# 先確保 PATH 包含 pi
+export PATH="$HOME/.local/bin:$HOME/.pi/bin:$(dirname "$PI_BIN"):$PATH"
+
+# 使用 timeout 防止卡死，並捕獲輸出
+set +e
+PI_TEST_OUTPUT=$(timeout 90 pi-local -p "用一句話介紹你自己，不超過20字" 2>&1)
+PI_EXIT_CODE=$?
+set -e
+
+if [[ $PI_EXIT_CODE -eq 0 ]] && [[ -n "$PI_TEST_OUTPUT" ]]; then
+    ok "pi-local 推理測試通過 (exit 0)"
+    echo "--- pi-local 輸出 ---"
+    echo "$PI_TEST_OUTPUT" | head -n 30
+    echo "---------------------"
+elif [[ $PI_EXIT_CODE -eq 124 ]]; then
+    warn "pi-local 測試超時 (90s)，可能模型載入較慢或首次啟動較久。請稍後手動重試: pi-local -p \"你好\""
+    # 不算致命失敗，因為 API 已通過
+else
+    warn "pi-local 測試失敗 (exit $PI_EXIT_CODE)，輸出:"
+    echo "$PI_TEST_OUTPUT" | head -n 100 >&2
+    # 如果 API 測試已通過，pi-local 失敗可能是設定問題，標記警告而非致命
+    if [[ "$VERIFICATION_FAILED" -eq 0 ]]; then
+        # 嘗試直接用 pi 而非 wrapper 再測一次
+        info "嘗試直接執行 pi 指令..."
+        set +e
+        PI_DIRECT_OUTPUT=$(timeout 90 pi --provider "$PROVIDER_ID" --model "$MODEL_ID" -p "你好" 2>&1)
+        PI_DIRECT_CODE=$?
+        set -e
+        if [[ $PI_DIRECT_CODE -eq 0 ]]; then
+            ok "pi 直接呼叫測試通過"
+        else
+            warn "pi 直接呼叫也失敗 (exit $PI_DIRECT_CODE): $PI_DIRECT_OUTPUT"
+            VERIFICATION_FAILED=1
+        fi
+    else
+        VERIFICATION_FAILED=1
+    fi
+fi
+
+# ==============================================================================
+# 總結
+# ==============================================================================
+echo ""
+echo "================================================================"
+if [[ "$VERIFICATION_FAILED" -eq 0 ]]; then
+    echo -e "\033[32m✅ 安裝完成且驗證通過！\033[0m"
+else
+    echo -e "\033[33m⚠️  安裝完成，但部分驗證未通過，請查看上方警告\033[0m"
+fi
+echo "================================================================"
+echo "- 模型: $REPO"
+echo "  主模型: $MODEL_PATH"
+if [[ -n "$MTP_PATH" ]]; then echo "  MTP 模型: $MTP_PATH"; fi
+echo "- llama-server: http://127.0.0.1:$LLAMA_PORT (log: ~/llama-server.log, pid: ~/llama-server.pid)"
+echo "- pi: $PI_BIN ($(pi --version 2>&1 | head -n 1))"
+echo "- pi 配置: $PI_MODELS_JSON (provider: $PROVIDER_ID, model: $MODEL_ID)"
+echo "- 快速使用:"
+echo "    source ~/.bashrc   # 或 source ~/.zshrc"
+echo "    pi-local -p \"你好\""
+echo "    pi --provider $PROVIDER_ID --model $MODEL_ID -p \"你好\""
+echo "- 停止伺服器: kill \$(cat ~/llama-server.pid)"
+echo "- 重啟伺服器: bash ~/llama-cpp/bin/  # 或重新執行本腳本"
+echo "- 重新驗證: curl http://127.0.0.1:$LLAMA_PORT/v1/models && pi-local -p \"測試\""
+echo "================================================================"
+
+if [[ "$VERIFICATION_FAILED" -ne 0 ]]; then
+    echo "請檢查日誌："
+    echo "  - llama-server: ~/llama-server.log (tail -n 100 ~/llama-server.log)"
+    echo "  - pi 安裝: /tmp/pi-install.log"
+    echo "  - 模型: $MODEL_DIR/"
+    exit 1
+fi
+
+exit 0