Spaces:

HelloSun
/

testzerogpu1

Paused

Copy from HelloSun/testzerogpu

1949064 verified about 9 hours ago

20.8 kB

#!/usr/bin/env bash
set -euo pipefail

# ==============================================================================
# installpi.sh - 一键全自动安装 pi + llama.cpp (源码编译) + Qwen2.5 模型
# 特性：无 sudo、无互动、全自动、可断点续传、安装后自动验证
# 模型：unsloth/GLM-4.7-Flash-GGUF (GLM-4.7-Flash-Q4_K_M.gguf)
# 启动：taskset -c 0-14 llama-server ... -t 15 --jinja --load-mode mlock -c 65536 (64K上下文)
# ==============================================================================

HOME_DIR="$HOME"
# 脚本一开始就把之前安装的 node / pi / llama 路径放进 PATH，避免每次重复装 node
export PATH="${XDG_DATA_HOME:-$HOME/.local/share}/pi-node/current/bin:$HOME/.local/bin:$HOME/.pi/bin:$PATH"
LLAMA_SRC_DIR="$HOME_DIR/llama.cpp"          # 源码目录
LLAMA_BIN="$LLAMA_SRC_DIR/build/bin"         # 编译后二进制目录
MODEL_DIR="$HOME_DIR/models"
PI_MODELS_JSON="$HOME_DIR/.pi/agent/models.json"
LLAMA_PORT=8080
REPO="unsloth/GLM-4.7-Flash-GGUF"
MODEL_FILE="GLM-4.7-Flash-Q4_K_M.gguf"
MODEL_ID="glm-4.7-flash-q4_k_m"
PROVIDER_ID="llama-cpp-local"
CONTEXT_SIZE=65536                          # 上下文 64K

# 基础输出函数
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; }

# 确保必要目录
mkdir -p "$LLAMA_SRC_DIR" "$MODEL_DIR" "$(dirname "$PI_MODELS_JSON")" "$HOME_DIR/.local/bin"

# 检查必要工具（新增编译依赖）
for cmd in curl python3 tar git cmake g++; do
    command -v "$cmd" >/dev/null 2>&1 || error "缺少必要工具: $cmd，请先安装"
done

# ==============================================================================
# 0. 确保 Node.js >=22.19.0 (pi 依赖)
# ==============================================================================
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 已满足: $(node --version)"
            return 0
        else
            warn "Node.js 版本过旧: $(node --version)，将自动安装新版..."
        fi
    else
        info "未检测到 Node.js，将自动安装..."
    fi

    info "自动安装 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"

    OS="$(uname -s)"
    ARCH="$(uname -m)"
    case "$OS" in Linux) PLATFORM="linux" ;; Darwin) PLATFORM="darwin" ;; *) error "不支持的 OS: $OS" ;; esac
    case "$ARCH" in x86_64|amd64) NODE_ARCH="x64" ;; aarch64|arm64) NODE_ARCH="arm64" ;; *) error "不支持的架构: $ARCH" ;; esac

    info "解析 Node.js 下载链接 ($PLATFORM-$NODE_ARCH)..."
    curl -fsSL --retry 3 --retry-delay 5 "$NODE_DIST_BASE/SHASUMS256.txt" -o "$NODE_TMP_DIR/SHASUMS256.txt" || error "下载 SHASUMS 失败"

    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 "找不到对应平台的 Node.js 二进制"

    info "下载 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 下载失败"

    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 校验失败，继续安装..."
    fi

    if ! command -v xz >/dev/null 2>&1; then
        warn "未找到 xz，将尝试使用 tar 自动解压 (若失败请手动安装 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 "解压 Node.js 失败"
    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"
    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 安装完成: $(node --version) at $(command -v node)"
    else
        error "Node.js 安装后仍无法找到，请检查 PATH"
    fi
}
ensure_node

command -v npm >/dev/null 2>&1 || error "npm 仍不可用，Node 安装可能失败"
info "npm 版本: $(npm --version)"

# ==============================================================================
# 1. 编译 llama.cpp (从源码)
# ==============================================================================
info "1) 编译 llama.cpp (源码)..."

if [[ "${LLAMA_FORCE_REBUILD:-0}" == "1" ]]; then
    info "LLAMA_FORCE_REBUILD=1，清理旧 build 目录强制重编..."
    rm -rf "$LLAMA_SRC_DIR/build"
fi

if [[ ! -x "$LLAMA_BIN/llama-server" ]]; then
    info "克隆 llama.cpp 仓库 (深度1)..."
    if [[ -d "$LLAMA_SRC_DIR/.git" ]]; then
        info "源码目录已存在，更新..."
        cd "$LLAMA_SRC_DIR"
        git pull || warn "git pull 失败，继续使用现有源码"
    else
        git clone --depth 1 https://github.com/ggml-org/llama.cpp "$LLAMA_SRC_DIR"
    fi

    cd "$LLAMA_SRC_DIR"
    info "执行 cmake 配置 (纯 CPU，不开 NATIVE/DL，避免 MoE/多变体后端在纯 CPU 上崩)..."
    rm -rf build
    cmake -B build \
        -DGGML_NATIVE=OFF \
        -DGGML_CPU=ON \
        -DGGML_BACKEND_DL=OFF \
        -DLLAMA_BUILD_TESTS=OFF \
        -DCMAKE_BUILD_TYPE=Release

    info "开始编译 (使用 -j16)..."
    cmake --build build --config Release -j 16

    # 验证编译产物
    if [[ -x "$LLAMA_BIN/llama-server" ]]; then
        ok "llama.cpp 编译完成: $LLAMA_BIN/llama-server"
    else
        error "编译失败，找不到 llama-server"
    fi
else
    ok "llama-server 已存在，跳过编译 (如需重编: LLAMA_FORCE_REBUILD=1 ./installpi.sh)"
fi

# 更新 PATH
export PATH="$LLAMA_BIN:$HOME_DIR/.local/bin:$PATH"
export LD_LIBRARY_PATH="$LLAMA_BIN:${LD_LIBRARY_PATH:-}"
for rc in "$HOME/.bashrc" "$HOME/.zshrc"; do
    touch "$rc" 2>/dev/null || true
    grep -q "llama.cpp/build/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. 下载模型 (直接使用指定 URL)
# ==============================================================================
info "2) 下载模型 ${REPO} -> ${MODEL_FILE} ..."

MODEL_URL="https://huggingface.co/${REPO}/resolve/main/${MODEL_FILE}"
MODEL_PATH="$MODEL_DIR/${MODEL_FILE}"

if [[ ! -f "$MODEL_PATH" ]]; then
    info "开始下载模型 (支持断点续传)..."
    curl -L -C - --retry 3 --retry-delay 5 --progress-bar "$MODEL_URL" -o "$MODEL_PATH" || error "模型下载失败"
    ok "模型下载完成: $MODEL_PATH"
else
    SIZE=$(stat -c%s "$MODEL_PATH" 2>/dev/null || stat -f%z "$MODEL_PATH" 2>/dev/null || echo 0)
    if [[ "$SIZE" -lt 1000000000 ]]; then
        warn "模型文件过小 (${SIZE} bytes)，可能下载不完整，将尝试续传..."
        curl -L -C - --retry 3 --retry-delay 5 --progress-bar "$MODEL_URL" -o "$MODEL_PATH" || error "续传失败"
    fi
    ok "模型已存在，跳过下载: $MODEL_PATH ($SIZE bytes)"
fi

# ==============================================================================
# 3. 启动 llama-server (使用 taskset 和指定参数，上下文 64K)
# ==============================================================================
info "3) 启动 llama-server (port $LLAMA_PORT, taskset 0-14, context $CONTEXT_SIZE)..."

# 清理旧进程
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 "停止旧 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

# 按实际 CPU 数决定 taskset 范围 / 线程数 (避免机器核数不足时 taskset 报错)
NPROC=$(nproc 2>/dev/null || echo 16)
if [[ "$NPROC" -ge 15 ]]; then
    TASKSET_ARGS=(-c 0-14)
    NTHREADS=15
else
    TASKSET_ARGS=(-c "0-$((NPROC-1))")
    NTHREADS="$NPROC"
fi
if ! taskset "${TASKSET_ARGS[@]}" true 2>/dev/null; then
    warn "taskset ${TASKSET_ARGS[*]} 不可用，改用不绑核启动"
    TASKSET_ARGS=()
fi

# 构建命令（prlimit 限制内存 32GB；LD_LIBRARY_PATH 已在前面 export，prlimit 会继承）
CMD=(
    taskset "${TASKSET_ARGS[@]}"
    prlimit --as=34359738368
    "$LLAMA_BIN/llama-server"
    -m "$MODEL_PATH"
    --alias "$MODEL_ID"
    -t "$NTHREADS"
    --jinja
    --load-mode mlock
    -c "$CONTEXT_SIZE"
    --host 127.0.0.1
    --port "$LLAMA_PORT"
)
# taskset 参数为空时去掉 taskset 本体
if [[ "${#TASKSET_ARGS[@]}" -eq 0 ]]; then
    CMD=(prlimit --as=34359738368 "$LLAMA_BIN/llama-server" -m "$MODEL_PATH" --alias "$MODEL_ID" -t "$NTHREADS" --jinja --load-mode mlock -c "$CONTEXT_SIZE" --host 127.0.0.1 --port "$LLAMA_PORT")
fi

info "启动命令: ${CMD[*]}"
touch "$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")"

# 等待服务器就绪 (最长 5 分钟)
info "等待 llama-server 就绪 (最多 5 分钟)..."
READY=0
for i in {1..60}; do
    if curl -sf "http://127.0.0.1:${LLAMA_PORT}/health" >/dev/null 2>&1; then
        ok "llama-server 已就绪 (http://127.0.0.1:${LLAMA_PORT})"
        READY=1
        break
    fi
    if curl -sf "http://127.0.0.1:${LLAMA_PORT}/v1/models" >/dev/null 2>&1; then
        ok "llama-server 已就绪 (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 启动失败，请查看 ~/llama-server.log (最后 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 分钟内未就绪，请查看 ~/llama-server.log"
fi

# ==============================================================================
# 4. 安装 pi agent (全自动，无互动)
# ==============================================================================
info "4) 安装 pi agent (全自动)..."

NPM_PREFIX=$(npm prefix -g 2>/dev/null || npm config get prefix 2>/dev/null || echo "$HOME/.local")
info "npm 全局前缀: $NPM_PREFIX"

# 确保 PATH 包含本地 bin 目录
export PATH="$HOME/.local/bin:$HOME/.pi/bin:$NPM_PREFIX/bin:$PATH"

NEED_INSTALL=1
if command -v pi >/dev/null 2>&1; then
    ok "pi 已安装: $(pi --version 2>&1 | head -n 1)"
    NEED_INSTALL=0
    if [[ "${PI_FORCE_REINSTALL:-0}" == "1" ]]; then
        info "PI_FORCE_REINSTALL=1，强制重新安装..."
        NEED_INSTALL=1
    fi
fi

if [[ "$NEED_INSTALL" -eq 1 ]]; then
    info "安装 pi (@earendil-works/pi-coding-agent)..."
    PI_PACKAGE="@earendil-works/pi-coding-agent"
    
    # 优先安装到 ~/.local (无需 sudo)
    INSTALL_PREFIX="$HOME/.local"
    info "安装到 $INSTALL_PREFIX (无需 sudo)"
    
    set +e
    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]}
    set -e

    if [[ "$NPM_STATUS" -ne 0 ]]; then
        warn "npm 安装 pi 失败 (code $NPM_STATUS)，尝试备用方法: curl 官方脚本..."
        curl -fsSL https://pi.dev/install.sh | bash -s -- 2>&1 | tee -a /tmp/pi-install.log || true
    fi

    export PATH="$HOME/.local/bin:$HOME/.pi/bin:$NPM_PREFIX/bin:$PATH"
    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

PI_BIN=$(command -v pi 2>/dev/null || echo "")
if [[ -z "$PI_BIN" ]]; then
    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 安装后仍找不到执行档，请查看 /tmp/pi-install.log"
fi

ok "pi 可用: $PI_BIN ($(pi --version 2>&1 | head -n 1))"

for rc in "$HOME/.bashrc" "$HOME/.zshrc" "$HOME/.profile"; do
    touch "$rc" 2>/dev/null || true
    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\"" >> "$rc"
    fi
    grep -qF ".local/bin" "$rc" 2>/dev/null || echo "export PATH=\"\$HOME/.local/bin:\$PATH\"" >> "$rc"
done

# ==============================================================================
# 5. 配置 pi 使用本地模型 (上下文调至极限)
# ==============================================================================
info "5) 配置 pi models.json ($PI_MODELS_JSON)..."

mkdir -p "$(dirname "$PI_MODELS_JSON")"

python3 - "$PI_MODELS_JSON" "$LLAMA_PORT" "$MODEL_ID" "$PROVIDER_ID" "$CONTEXT_SIZE" << 'PY'
import json, os, sys
path, port, model_id, provider, ctx_size = sys.argv[1], sys.argv[2], sys.argv[3], sys.argv[4], int(sys.argv[5])
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] 无法解析旧 models.json，将覆盖: {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": "GLM-4.7-Flash Q4_K_M (Local)",
        "reasoning": True,
        "input": ["text"],
        "contextWindow": ctx_size,
        "maxTokens": ctx_size,
        "cost": {"input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0}
    }]
}
with open(path, "w") as f:
    json.dump(cfg, f, indent=2)
print(f"[OK] 已写入 {path}")
print(json.dumps(cfg, indent=2))
PY

ok "pi 配置完成"

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"

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. 安装后验证
# ==============================================================================
info "6) 安装后验证..."

VERIFICATION_FAILED=0

# 6.1 检查 llama-server 健康
info "[验证 1/4] 检查 llama-server 健康状态..."
if curl -sf "http://127.0.0.1:${LLAMA_PORT}/health" >/dev/null 2>&1; then
    ok "llama-server 健康检查通过"
elif curl -sf "http://127.0.0.1:${LLAMA_PORT}/v1/models" >/dev/null 2>&1; then
    ok "llama-server (via /v1/models) 健康检查通过"
else
    warn "llama-server 健康检查失败"
    VERIFICATION_FAILED=1
fi

# 6.2 测试 OpenAI API 推理
info "[验证 2/4] 测试 OpenAI 兼容 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": "用一句话介绍你自己，不要超过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 推理测试通过"
    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 失败: $API_TEST_RESPONSE"
    VERIFICATION_FAILED=1
else
    warn "API 推理回应异常，可能是模型仍在加载中..."
    echo "$API_TEST_RESPONSE" | head -n 50 >&2
fi

# 6.3 检查 pi 可用
info "[验证 3/4] 检查 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 执行失败"
    VERIFICATION_FAILED=1
fi

if pi --help >/dev/null 2>&1; then
    ok "pi --help 正常"
fi

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 配置正确"
else
    warn "pi models.json 配置检查失败"
    cat "$PI_MODELS_JSON" 2>/dev/null || echo "文件不存在"
    VERIFICATION_FAILED=1
fi

if [[ -x "$WRAPPER" ]]; then
    ok "pi-local wrapper 存在且可执行"
else
    warn "pi-local wrapper 缺失"
    VERIFICATION_FAILED=1
fi

# 6.4 测试 pi-local 实际可用
info "[验证 4/4] 测试 pi-local 实际推理 (超时 120 秒)..."
export PATH="$HOME/.local/bin:$HOME/.pi/bin:$(dirname "$PI_BIN"):$PATH"
hash -r 2>/dev/null || true

set +e
PI_TEST_OUTPUT=$(timeout 120 pi --provider "$PROVIDER_ID" --model "$MODEL_ID" -p "用一句话介绍你自己，不超过20字" 2>&1)
PI_EXIT_CODE=$?
set -e

if [[ $PI_EXIT_CODE -eq 0 ]] && [[ -n "$PI_TEST_OUTPUT" ]]; then
    ok "pi 推理测试通过 (exit 0)"
    echo "--- pi 输出 ---"
    echo "$PI_TEST_OUTPUT" | head -n 30
    echo "----------------"
elif [[ $PI_EXIT_CODE -eq 124 ]]; then
    warn "pi 测试超时 (120s)，可能模型加载较慢或首次启动较久。请稍后手动重试: pi --provider $PROVIDER_ID --model $MODEL_ID -p \"你好\""
else
    warn "pi 测试失败 (exit $PI_EXIT_CODE)，输出:"
    echo "$PI_TEST_OUTPUT" | head -n 100 >&2
    VERIFICATION_FAILED=1
fi

# ==============================================================================
# 总结
# ==============================================================================
echo ""
echo "================================================================"
if [[ "$VERIFICATION_FAILED" -eq 0 ]]; then
    echo -e "\033[32m✅ 安装完成且验证通过！\033[0m"
else
    echo -e "\033[33m⚠️  安装完成，但部分验证未通过，请查看上方警告\033[0m"
fi
echo "================================================================"
echo "- 模型: $REPO ($MODEL_FILE)"
echo "  模型文件: $MODEL_PATH"
echo "- 内存限制: 32GB"
echo "- llama-server (编译自源码): $LLAMA_BIN/llama-server"
echo "  启动参数: taskset -c 0-14 ... -t 15 --jinja --load-mode mlock -c $CONTEXT_SIZE --host 127.0.0.1 --port $LLAMA_PORT"
echo "- 服务: http://127.0.0.1:$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, context: $CONTEXT_SIZE)"
echo "- 快速使用:"
echo "    source ~/.bashrc   # 或 source ~/.zshrc"
echo "    pi-local -p \"你好\""
echo "    pi --provider $PROVIDER_ID --model $MODEL_ID -p \"你好\""
echo "- 停止服务器: kill \$(cat ~/llama-server.pid)"
echo "- 重新验证: curl http://127.0.0.1:$LLAMA_PORT/v1/models && pi-local -p \"测试\""
echo "================================================================"

if [[ "$VERIFICATION_FAILED" -ne 0 ]]; then
    echo "请检查日志："
    echo "  - llama-server: ~/llama-server.log (tail -n 100 ~/llama-server.log)"
    echo "  - pi 安装: /tmp/pi-install.log"
    echo "  - 模型: $MODEL_DIR/"
    exit 1
fi

exit 0

	#!/usr/bin/env bash
	set -euo pipefail

	# ==============================================================================
	# installpi.sh - 一键全自动安装 pi + llama.cpp (源码编译) + Qwen2.5 模型
	# 特性：无 sudo、无互动、全自动、可断点续传、安装后自动验证
	# 模型：unsloth/GLM-4.7-Flash-GGUF (GLM-4.7-Flash-Q4_K_M.gguf)
	# 启动：taskset -c 0-14 llama-server ... -t 15 --jinja --load-mode mlock -c 65536 (64K上下文)
	# ==============================================================================

	HOME_DIR="$HOME"
	# 脚本一开始就把之前安装的 node / pi / llama 路径放进 PATH，避免每次重复装 node
	export PATH="${XDG_DATA_HOME:-$HOME/.local/share}/pi-node/current/bin:$HOME/.local/bin:$HOME/.pi/bin:$PATH"
	LLAMA_SRC_DIR="$HOME_DIR/llama.cpp" # 源码目录
	LLAMA_BIN="$LLAMA_SRC_DIR/build/bin" # 编译后二进制目录
	MODEL_DIR="$HOME_DIR/models"
	PI_MODELS_JSON="$HOME_DIR/.pi/agent/models.json"
	LLAMA_PORT=8080
	REPO="unsloth/GLM-4.7-Flash-GGUF"
	MODEL_FILE="GLM-4.7-Flash-Q4_K_M.gguf"
	MODEL_ID="glm-4.7-flash-q4_k_m"
	PROVIDER_ID="llama-cpp-local"
	CONTEXT_SIZE=65536 # 上下文 64K

	# 基础输出函数
	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; }

	# 确保必要目录
	mkdir -p "$LLAMA_SRC_DIR" "$MODEL_DIR" "$(dirname "$PI_MODELS_JSON")" "$HOME_DIR/.local/bin"

	# 检查必要工具（新增编译依赖）
	for cmd in curl python3 tar git cmake g++; do
	command -v "$cmd" >/dev/null 2>&1 \|\| error "缺少必要工具: $cmd，请先安装"
	done

	# ==============================================================================
	# 0. 确保 Node.js >=22.19.0 (pi 依赖)
	# ==============================================================================
	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 已满足: $(node --version)"
	return 0
	else
	warn "Node.js 版本过旧: $(node --version)，将自动安装新版..."
	fi
	else
	info "未检测到 Node.js，将自动安装..."
	fi

	info "自动安装 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"

	OS="$(uname -s)"
	ARCH="$(uname -m)"
	case "$OS" in Linux) PLATFORM="linux" ;; Darwin) PLATFORM="darwin" ;; *) error "不支持的 OS: $OS" ;; esac
	case "$ARCH" in x86_64\|amd64) NODE_ARCH="x64" ;; aarch64\|arm64) NODE_ARCH="arm64" ;; *) error "不支持的架构: $ARCH" ;; esac

	info "解析 Node.js 下载链接 ($PLATFORM-$NODE_ARCH)..."
	curl -fsSL --retry 3 --retry-delay 5 "$NODE_DIST_BASE/SHASUMS256.txt" -o "$NODE_TMP_DIR/SHASUMS256.txt" \|\| error "下载 SHASUMS 失败"

	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 "找不到对应平台的 Node.js 二进制"

	info "下载 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 下载失败"

	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 校验失败，继续安装..."
	fi

	if ! command -v xz >/dev/null 2>&1; then
	warn "未找到 xz，将尝试使用 tar 自动解压 (若失败请手动安装 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 "解压 Node.js 失败"
	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"
	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 安装完成: $(node --version) at $(command -v node)"
	else
	error "Node.js 安装后仍无法找到，请检查 PATH"
	fi
	}
	ensure_node

	command -v npm >/dev/null 2>&1 \|\| error "npm 仍不可用，Node 安装可能失败"
	info "npm 版本: $(npm --version)"

	# ==============================================================================
	# 1. 编译 llama.cpp (从源码)
	# ==============================================================================
	info "1) 编译 llama.cpp (源码)..."

	if [[ "${LLAMA_FORCE_REBUILD:-0}" == "1" ]]; then
	info "LLAMA_FORCE_REBUILD=1，清理旧 build 目录强制重编..."
	rm -rf "$LLAMA_SRC_DIR/build"
	fi

	if [[ ! -x "$LLAMA_BIN/llama-server" ]]; then
	info "克隆 llama.cpp 仓库 (深度1)..."
	if [[ -d "$LLAMA_SRC_DIR/.git" ]]; then
	info "源码目录已存在，更新..."
	cd "$LLAMA_SRC_DIR"
	git pull \|\| warn "git pull 失败，继续使用现有源码"
	else
	git clone --depth 1 https://github.com/ggml-org/llama.cpp "$LLAMA_SRC_DIR"
	fi

	cd "$LLAMA_SRC_DIR"
	info "执行 cmake 配置 (纯 CPU，不开 NATIVE/DL，避免 MoE/多变体后端在纯 CPU 上崩)..."
	rm -rf build
	cmake -B build \
	-DGGML_NATIVE=OFF \
	-DGGML_CPU=ON \
	-DGGML_BACKEND_DL=OFF \
	-DLLAMA_BUILD_TESTS=OFF \
	-DCMAKE_BUILD_TYPE=Release

	info "开始编译 (使用 -j16)..."
	cmake --build build --config Release -j 16

	# 验证编译产物
	if [[ -x "$LLAMA_BIN/llama-server" ]]; then
	ok "llama.cpp 编译完成: $LLAMA_BIN/llama-server"
	else
	error "编译失败，找不到 llama-server"
	fi
	else
	ok "llama-server 已存在，跳过编译 (如需重编: LLAMA_FORCE_REBUILD=1 ./installpi.sh)"
	fi

	# 更新 PATH
	export PATH="$LLAMA_BIN:$HOME_DIR/.local/bin:$PATH"
	export LD_LIBRARY_PATH="$LLAMA_BIN:${LD_LIBRARY_PATH:-}"
	for rc in "$HOME/.bashrc" "$HOME/.zshrc"; do
	touch "$rc" 2>/dev/null \|\| true
	grep -q "llama.cpp/build/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. 下载模型 (直接使用指定 URL)
	# ==============================================================================
	info "2) 下载模型 ${REPO} -> ${MODEL_FILE} ..."

	MODEL_URL="https://huggingface.co/${REPO}/resolve/main/${MODEL_FILE}"
	MODEL_PATH="$MODEL_DIR/${MODEL_FILE}"

	if [[ ! -f "$MODEL_PATH" ]]; then
	info "开始下载模型 (支持断点续传)..."
	curl -L -C - --retry 3 --retry-delay 5 --progress-bar "$MODEL_URL" -o "$MODEL_PATH" \|\| error "模型下载失败"
	ok "模型下载完成: $MODEL_PATH"
	else
	SIZE=$(stat -c%s "$MODEL_PATH" 2>/dev/null \|\| stat -f%z "$MODEL_PATH" 2>/dev/null \|\| echo 0)
	if [[ "$SIZE" -lt 1000000000 ]]; then
	warn "模型文件过小 (${SIZE} bytes)，可能下载不完整，将尝试续传..."
	curl -L -C - --retry 3 --retry-delay 5 --progress-bar "$MODEL_URL" -o "$MODEL_PATH" \|\| error "续传失败"
	fi
	ok "模型已存在，跳过下载: $MODEL_PATH ($SIZE bytes)"
	fi

	# ==============================================================================
	# 3. 启动 llama-server (使用 taskset 和指定参数，上下文 64K)
	# ==============================================================================
	info "3) 启动 llama-server (port $LLAMA_PORT, taskset 0-14, context $CONTEXT_SIZE)..."

	# 清理旧进程
	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 "停止旧 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

	# 按实际 CPU 数决定 taskset 范围 / 线程数 (避免机器核数不足时 taskset 报错)
	NPROC=$(nproc 2>/dev/null \|\| echo 16)
	if [[ "$NPROC" -ge 15 ]]; then
	TASKSET_ARGS=(-c 0-14)
	NTHREADS=15
	else
	TASKSET_ARGS=(-c "0-$((NPROC-1))")
	NTHREADS="$NPROC"
	fi
	if ! taskset "${TASKSET_ARGS[@]}" true 2>/dev/null; then
	warn "taskset ${TASKSET_ARGS[*]} 不可用，改用不绑核启动"
	TASKSET_ARGS=()
	fi

	# 构建命令（prlimit 限制内存 32GB；LD_LIBRARY_PATH 已在前面 export，prlimit 会继承）
	CMD=(
	taskset "${TASKSET_ARGS[@]}"
	prlimit --as=34359738368
	"$LLAMA_BIN/llama-server"
	-m "$MODEL_PATH"
	--alias "$MODEL_ID"
	-t "$NTHREADS"
	--jinja
	--load-mode mlock
	-c "$CONTEXT_SIZE"
	--host 127.0.0.1
	--port "$LLAMA_PORT"
	)
	# taskset 参数为空时去掉 taskset 本体
	if [[ "${#TASKSET_ARGS[@]}" -eq 0 ]]; then
	CMD=(prlimit --as=34359738368 "$LLAMA_BIN/llama-server" -m "$MODEL_PATH" --alias "$MODEL_ID" -t "$NTHREADS" --jinja --load-mode mlock -c "$CONTEXT_SIZE" --host 127.0.0.1 --port "$LLAMA_PORT")
	fi

	info "启动命令: ${CMD[*]}"
	touch "$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")"

	# 等待服务器就绪 (最长 5 分钟)
	info "等待 llama-server 就绪 (最多 5 分钟)..."
	READY=0
	for i in {1..60}; do
	if curl -sf "http://127.0.0.1:${LLAMA_PORT}/health" >/dev/null 2>&1; then
	ok "llama-server 已就绪 (http://127.0.0.1:${LLAMA_PORT})"
	READY=1
	break
	fi
	if curl -sf "http://127.0.0.1:${LLAMA_PORT}/v1/models" >/dev/null 2>&1; then
	ok "llama-server 已就绪 (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 启动失败，请查看 ~/llama-server.log (最后 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 分钟内未就绪，请查看 ~/llama-server.log"
	fi

	# ==============================================================================
	# 4. 安装 pi agent (全自动，无互动)
	# ==============================================================================
	info "4) 安装 pi agent (全自动)..."

	NPM_PREFIX=$(npm prefix -g 2>/dev/null \|\| npm config get prefix 2>/dev/null \|\| echo "$HOME/.local")
	info "npm 全局前缀: $NPM_PREFIX"

	# 确保 PATH 包含本地 bin 目录
	export PATH="$HOME/.local/bin:$HOME/.pi/bin:$NPM_PREFIX/bin:$PATH"

	NEED_INSTALL=1
	if command -v pi >/dev/null 2>&1; then
	ok "pi 已安装: $(pi --version 2>&1 \| head -n 1)"
	NEED_INSTALL=0
	if [[ "${PI_FORCE_REINSTALL:-0}" == "1" ]]; then
	info "PI_FORCE_REINSTALL=1，强制重新安装..."
	NEED_INSTALL=1
	fi
	fi

	if [[ "$NEED_INSTALL" -eq 1 ]]; then
	info "安装 pi (@earendil-works/pi-coding-agent)..."
	PI_PACKAGE="@earendil-works/pi-coding-agent"

	# 优先安装到 ~/.local (无需 sudo)
	INSTALL_PREFIX="$HOME/.local"
	info "安装到 $INSTALL_PREFIX (无需 sudo)"

	set +e
	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]}
	set -e

	if [[ "$NPM_STATUS" -ne 0 ]]; then
	warn "npm 安装 pi 失败 (code $NPM_STATUS)，尝试备用方法: curl 官方脚本..."
	curl -fsSL https://pi.dev/install.sh \| bash -s -- 2>&1 \| tee -a /tmp/pi-install.log \|\| true
	fi

	export PATH="$HOME/.local/bin:$HOME/.pi/bin:$NPM_PREFIX/bin:$PATH"
	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

	PI_BIN=$(command -v pi 2>/dev/null \|\| echo "")
	if [[ -z "$PI_BIN" ]]; then
	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 安装后仍找不到执行档，请查看 /tmp/pi-install.log"
	fi

	ok "pi 可用: $PI_BIN ($(pi --version 2>&1 \| head -n 1))"

	for rc in "$HOME/.bashrc" "$HOME/.zshrc" "$HOME/.profile"; do
	touch "$rc" 2>/dev/null \|\| true
	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\"" >> "$rc"
	fi
	grep -qF ".local/bin" "$rc" 2>/dev/null \|\| echo "export PATH=\"\$HOME/.local/bin:\$PATH\"" >> "$rc"
	done

	# ==============================================================================
	# 5. 配置 pi 使用本地模型 (上下文调至极限)
	# ==============================================================================
	info "5) 配置 pi models.json ($PI_MODELS_JSON)..."

	mkdir -p "$(dirname "$PI_MODELS_JSON")"

	python3 - "$PI_MODELS_JSON" "$LLAMA_PORT" "$MODEL_ID" "$PROVIDER_ID" "$CONTEXT_SIZE" << 'PY'
	import json, os, sys
	path, port, model_id, provider, ctx_size = sys.argv[1], sys.argv[2], sys.argv[3], sys.argv[4], int(sys.argv[5])
	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] 无法解析旧 models.json，将覆盖: {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": "GLM-4.7-Flash Q4_K_M (Local)",
	"reasoning": True,
	"input": ["text"],
	"contextWindow": ctx_size,
	"maxTokens": ctx_size,
	"cost": {"input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0}
	}]
	}
	with open(path, "w") as f:
	json.dump(cfg, f, indent=2)
	print(f"[OK] 已写入 {path}")
	print(json.dumps(cfg, indent=2))
	PY

	ok "pi 配置完成"

	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"

	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. 安装后验证
	# ==============================================================================
	info "6) 安装后验证..."

	VERIFICATION_FAILED=0

	# 6.1 检查 llama-server 健康
	info "[验证 1/4] 检查 llama-server 健康状态..."
	if curl -sf "http://127.0.0.1:${LLAMA_PORT}/health" >/dev/null 2>&1; then
	ok "llama-server 健康检查通过"
	elif curl -sf "http://127.0.0.1:${LLAMA_PORT}/v1/models" >/dev/null 2>&1; then
	ok "llama-server (via /v1/models) 健康检查通过"
	else
	warn "llama-server 健康检查失败"
	VERIFICATION_FAILED=1
	fi

	# 6.2 测试 OpenAI API 推理
	info "[验证 2/4] 测试 OpenAI 兼容 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": "用一句话介绍你自己，不要超过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 推理测试通过"
	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 失败: $API_TEST_RESPONSE"
	VERIFICATION_FAILED=1
	else
	warn "API 推理回应异常，可能是模型仍在加载中..."
	echo "$API_TEST_RESPONSE" \| head -n 50 >&2
	fi

	# 6.3 检查 pi 可用
	info "[验证 3/4] 检查 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 执行失败"
	VERIFICATION_FAILED=1
	fi

	if pi --help >/dev/null 2>&1; then
	ok "pi --help 正常"
	fi

	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 配置正确"
	else
	warn "pi models.json 配置检查失败"
	cat "$PI_MODELS_JSON" 2>/dev/null \|\| echo "文件不存在"
	VERIFICATION_FAILED=1
	fi

	if [[ -x "$WRAPPER" ]]; then
	ok "pi-local wrapper 存在且可执行"
	else
	warn "pi-local wrapper 缺失"
	VERIFICATION_FAILED=1
	fi

	# 6.4 测试 pi-local 实际可用
	info "[验证 4/4] 测试 pi-local 实际推理 (超时 120 秒)..."
	export PATH="$HOME/.local/bin:$HOME/.pi/bin:$(dirname "$PI_BIN"):$PATH"
	hash -r 2>/dev/null \|\| true

	set +e
	PI_TEST_OUTPUT=$(timeout 120 pi --provider "$PROVIDER_ID" --model "$MODEL_ID" -p "用一句话介绍你自己，不超过20字" 2>&1)
	PI_EXIT_CODE=$?
	set -e

	if [[ $PI_EXIT_CODE -eq 0 ]] && [[ -n "$PI_TEST_OUTPUT" ]]; then
	ok "pi 推理测试通过 (exit 0)"
	echo "--- pi 输出 ---"
	echo "$PI_TEST_OUTPUT" \| head -n 30
	echo "----------------"
	elif [[ $PI_EXIT_CODE -eq 124 ]]; then
	warn "pi 测试超时 (120s)，可能模型加载较慢或首次启动较久。请稍后手动重试: pi --provider $PROVIDER_ID --model $MODEL_ID -p \"你好\""
	else
	warn "pi 测试失败 (exit $PI_EXIT_CODE)，输出:"
	echo "$PI_TEST_OUTPUT" \| head -n 100 >&2
	VERIFICATION_FAILED=1
	fi

	# ==============================================================================
	# 总结
	# ==============================================================================
	echo ""
	echo "================================================================"
	if [[ "$VERIFICATION_FAILED" -eq 0 ]]; then
	echo -e "\033[32m✅ 安装完成且验证通过！\033[0m"
	else
	echo -e "\033[33m⚠️ 安装完成，但部分验证未通过，请查看上方警告\033[0m"
	fi
	echo "================================================================"
	echo "- 模型: $REPO ($MODEL_FILE)"
	echo " 模型文件: $MODEL_PATH"
	echo "- 内存限制: 32GB"
	echo "- llama-server (编译自源码): $LLAMA_BIN/llama-server"
	echo " 启动参数: taskset -c 0-14 ... -t 15 --jinja --load-mode mlock -c $CONTEXT_SIZE --host 127.0.0.1 --port $LLAMA_PORT"
	echo "- 服务: http://127.0.0.1:$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, context: $CONTEXT_SIZE)"
	echo "- 快速使用:"
	echo " source ~/.bashrc # 或 source ~/.zshrc"
	echo " pi-local -p \"你好\""
	echo " pi --provider $PROVIDER_ID --model $MODEL_ID -p \"你好\""
	echo "- 停止服务器: kill \$(cat ~/llama-server.pid)"
	echo "- 重新验证: curl http://127.0.0.1:$LLAMA_PORT/v1/models && pi-local -p \"测试\""
	echo "================================================================"

	if [[ "$VERIFICATION_FAILED" -ne 0 ]]; then
	echo "请检查日志："
	echo " - llama-server: ~/llama-server.log (tail -n 100 ~/llama-server.log)"
	echo " - pi 安装: /tmp/pi-install.log"
	echo " - 模型: $MODEL_DIR/"
	exit 1
	fi

	exit 0