From 24934bc182799fe8e505bb0d67469506014d8f94 Mon Sep 17 00:00:00 2001
From: qinyusen <qinyusen@gmail.com>
Date: Thu, 7 May 2026 01:32:13 +0800
Subject: [PATCH] feat: rewrite install_deps.sh with env isolation and add
 numpy to requirements
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- Complete rewrite of install_deps.sh (6-phase architecture):
  environment validation, uv-based venv isolation, CUDA auto-detection,
  idempotent native tool compilation, env.sh/run-gpu-tests generation
- Add numpy>=1.24 to requirements.txt to align with pyproject.toml
- Support --install-system-deps, --skip-pytorch, --rebuild, -y flags
- Use subshells for compilation to prevent CWD pollution
- Generate env.sh activation script and run-gpu-tests wrapper

🤖 Generated with [Qoder][https://qoder.com]
---
 install_deps.sh  | 815 ++++++++++++++++++++++++++++++++++++++---------
 requirements.txt |   1 +
 2 files changed, 666 insertions(+), 150 deletions(-)

diff --git a/install_deps.sh b/install_deps.sh
index c84b47b..a9659a3 100755
--- a/install_deps.sh
+++ b/install_deps.sh
@@ -1,234 +1,749 @@
 #!/usr/bin/env bash
-set -euo pipefail
+# =============================================================================
+# GPU Test Suite — 一键安装脚本（环境隔离版）
+# 支持: A100 / A800 / H100 / H200 / B200 / B300
+#
+# 功能:
+#   1. 环境校验（GPU、CUDA、Python、编译器等）
+#   2. 自动安装 uv 并创建隔离 Python 虚拟环境
+#   3. 自动检测 CUDA 版本并安装对应 PyTorch
+#   4. 编译 nvbandwidth / nccl-tests / gpu-burn
+#   5. 生成 env.sh 激活脚本和 run-gpu-tests 运行器
+#
+# 用法:
+#   sudo bash install_deps.sh                   # 标准安装
+#   sudo bash install_deps.sh --install-system-deps  # 同时安装系统包
+#   sudo bash install_deps.sh --skip-pytorch    # 跳过 PyTorch
+#   sudo bash install_deps.sh --rebuild         # 强制重新编译
+#   sudo bash install_deps.sh -y                # 非交互模式
+# =============================================================================
+set -uo pipefail
 
+# ─── 全局变量 ─────────────────────────────────────────────────────────────────
 INSTALL_DIR="${GPU_TOOLS_DIR:-/opt/gpu-test-tools}"
+PROJECT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 JOBS="${MAKE_JOBS:-$(nproc)}"
 VERBOSE="${VERBOSE:-0}"
 
+# 参数标志
+FLAG_INSTALL_SYS_DEPS=0
+FLAG_SKIP_PYTORCH=0
+FLAG_REBUILD=0
+FLAG_YES=0
+
+# 检测结果（全局）
+DETECTED_GPU=""
+DETECTED_DRIVER=""
+CUDA_VERSION=""
+CUDA_TAG=""
+PYTHON_BIN=""
+HAS_MPI=0
+HAS_NCCL_DEV=0
+
+# ─── 颜色和日志 ──────────────────────────────────────────────────────────────
 RED='\033[0;31m'
 GREEN='\033[0;32m'
 YELLOW='\033[1;33m'
 CYAN='\033[0;36m'
+BOLD='\033[1m'
 NC='\033[0m'
 
 log()  { echo -e "${CYAN}[INFO]${NC} $*"; }
 ok()   { echo -e "${GREEN}[ OK ]${NC} $*"; }
 warn() { echo -e "${YELLOW}[WARN]${NC} $*"; }
 fail() { echo -e "${RED}[FAIL]${NC} $*"; }
+die()  { echo -e "${RED}[FATAL]${NC} $*"; exit 1; }
+banner() { echo -e "\n${BOLD}${CYAN}══════ $* ══════${NC}\n"; }
 
-check_root() {
-    if [[ $EUID -ne 0 ]]; then
-        warn "Not running as root. Some installations may fail."
-        warn "Re-run with: sudo $0"
-    fi
+# 错误陷阱
+trap 'fail "脚本在第 $LINENO 行出错。设置 VERBOSE=1 查看详情。"' ERR
+
+# ─── 参数解析 ─────────────────────────────────────────────────────────────────
+parse_args() {
+    while [[ $# -gt 0 ]]; do
+        case "$1" in
+            --install-system-deps) FLAG_INSTALL_SYS_DEPS=1 ;;
+            --skip-pytorch)        FLAG_SKIP_PYTORCH=1 ;;
+            --rebuild)             FLAG_REBUILD=1 ;;
+            -y|--yes)              FLAG_YES=1 ;;
+            -h|--help)
+                echo "用法: $0 [选项]"
+                echo ""
+                echo "选项:"
+                echo "  --install-system-deps  自动安装缺失的系统包"
+                echo "  --skip-pytorch         跳过 PyTorch 安装"
+                echo "  --rebuild              强制重新编译原生工具"
+                echo "  -y, --yes              非交互模式"
+                echo "  -h, --help             显示此帮助"
+                echo ""
+                echo "环境变量:"
+                echo "  GPU_TOOLS_DIR          安装目录 (默认: /opt/gpu-test-tools)"
+                echo "  MAKE_JOBS              编译并行数 (默认: nproc)"
+                echo "  CUDA_HOME              CUDA 安装路径 (默认: /usr/local/cuda)"
+                exit 0
+                ;;
+            *) warn "未知参数: $1" ;;
+        esac
+        shift
+    done
 }
 
-detect_gpu() {
-    if ! command -v nvidia-smi &>/dev/null; then
-        fail "nvidia-smi not found. Install NVIDIA drivers first."
-        exit 1
+# ─── 阶段 0: 权限检查 ─────────────────────────────────────────────────────────
+check_permissions() {
+    local parent_dir
+    parent_dir="$(dirname "$INSTALL_DIR")"
+    if [[ ! -w "$parent_dir" ]] && [[ ! -d "$INSTALL_DIR" || ! -w "$INSTALL_DIR" ]]; then
+        die "无法写入 $INSTALL_DIR（请使用 sudo 或设置 GPU_TOOLS_DIR 到可写路径）"
     fi
-    local gpu_name
-    gpu_name=$(nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null | head -1)
-    log "Detected GPU: $gpu_name"
+    mkdir -p "$INSTALL_DIR"
+}
+
+# ─── 阶段 1: 环境校验 ─────────────────────────────────────────────────────────
+
+# 最低驱动版本表
+declare -A MIN_DRIVERS=(
+    ["A100"]="470" ["A800"]="470"
+    ["H100"]="535" ["H200"]="535"
+    ["B200"]="550" ["B300"]="550"
+)
+
+check_nvidia_smi() {
+    if ! command -v nvidia-smi &>/dev/null; then
+        fail "nvidia-smi 未找到"
+        echo "  → 请先安装 NVIDIA 驱动"
+        return 1
+    fi
+    ok "nvidia-smi 可用"
+    return 0
+}
+
+detect_gpu_and_driver() {
+    local smi_out
+    smi_out=$(nvidia-smi --query-gpu=name,driver_version --format=csv,noheader 2>/dev/null | head -1)
+    if [[ -z "$smi_out" ]]; then
+        warn "无法查询 GPU 信息"
+        return 1
+    fi
+
+    DETECTED_GPU=$(echo "$smi_out" | cut -d',' -f1 | xargs)
+    DETECTED_DRIVER=$(echo "$smi_out" | cut -d',' -f2 | xargs)
+
+    # 检查驱动版本
+    local gpu_key=""
+    for key in "${!MIN_DRIVERS[@]}"; do
+        if echo "$DETECTED_GPU" | grep -qi "$key"; then
+            gpu_key="$key"
+            break
+        fi
+    done
+
+    if [[ -n "$gpu_key" ]]; then
+        local min_drv="${MIN_DRIVERS[$gpu_key]}"
+        local drv_major="${DETECTED_DRIVER%%.*}"
+        if [[ "$drv_major" -lt "$min_drv" ]]; then
+            warn "驱动 $DETECTED_DRIVER < 最低要求 $min_drv（$gpu_key 需要）"
+        else
+            ok "GPU: $DETECTED_GPU | 驱动: $DETECTED_DRIVER (>= $min_drv)"
+        fi
+    else
+        ok "GPU: $DETECTED_GPU | 驱动: $DETECTED_DRIVER"
+    fi
+    return 0
+}
+
+detect_cuda_version() {
+    # 方式 1: nvcc（最可靠，代表 toolkit 确实安装了）
+    if command -v nvcc &>/dev/null; then
+        CUDA_VERSION=$(nvcc --version 2>/dev/null | grep -oP 'release \K[0-9]+\.[0-9]+')
+        if [[ -n "$CUDA_VERSION" ]]; then
+            ok "CUDA: $CUDA_VERSION (via nvcc)"
+            _map_cuda_tag
+            return 0
+        fi
+    fi
+
+    # 方式 2: nvidia-smi（驱动支持的最大 CUDA 版本，非 toolkit）
+    local smi_cuda
+    smi_cuda=$(nvidia-smi 2>/dev/null | grep -oP 'CUDA Version: \K[0-9]+\.[0-9]+')
+    if [[ -n "$smi_cuda" ]]; then
+        CUDA_VERSION="$smi_cuda"
+        warn "CUDA: $CUDA_VERSION (via nvidia-smi — 仅代表驱动能力，非已安装 toolkit)"
+        warn "  → 若编译失败，请安装 CUDA Toolkit: apt install cuda-toolkit-${CUDA_VERSION/./-}"
+        _map_cuda_tag
+        return 0
+    fi
+
+    # 方式 3: /usr/local/cuda
+    if [[ -f /usr/local/cuda/version.txt ]]; then
+        CUDA_VERSION=$(grep -oP '[0-9]+\.[0-9]+' /usr/local/cuda/version.txt | head -1)
+        if [[ -n "$CUDA_VERSION" ]]; then
+            ok "CUDA: $CUDA_VERSION (via /usr/local/cuda/version.txt)"
+            _map_cuda_tag
+            return 0
+        fi
+    fi
+
+    fail "无法检测 CUDA 版本"
+    echo "  → 请安装 CUDA Toolkit: https://developer.nvidia.com/cuda-downloads"
+    return 1
+}
+
+_map_cuda_tag() {
+    local major minor
+    major="${CUDA_VERSION%%.*}"
+    minor="${CUDA_VERSION#*.}"
+    minor="${minor%%.*}"
+
+    if [[ "$major" -eq 11 ]]; then
+        CUDA_TAG="cu118"
+    elif [[ "$major" -eq 12 ]]; then
+        if [[ "$minor" -le 1 ]]; then
+            CUDA_TAG="cu121"
+        elif [[ "$minor" -le 4 ]]; then
+            CUDA_TAG="cu124"
+        else
+            CUDA_TAG="cu128"
+        fi
+    else
+        CUDA_TAG="cu128"
+        warn "未知 CUDA $CUDA_VERSION，默认使用 cu128 索引"
+    fi
+    log "PyTorch wheel 索引: $CUDA_TAG"
+}
+
+check_python() {
+    local py_cmd=""
+    for cmd in python3.12 python3.11 python3.10 python3; do
+        if command -v "$cmd" &>/dev/null; then
+            local ver
+            ver=$("$cmd" -c "import sys; print(f'{sys.version_info.major}.{sys.version_info.minor}')" 2>/dev/null)
+            local py_major py_minor
+            py_major="${ver%%.*}"
+            py_minor="${ver#*.}"
+            if [[ "$py_major" -ge 3 ]] && [[ "$py_minor" -ge 10 ]]; then
+                py_cmd="$cmd"
+                break
+            fi
+        fi
+    done
+
+    if [[ -z "$py_cmd" ]]; then
+        fail "Python >= 3.10 未找到"
+        echo "  → apt install python3.11 python3.11-venv"
+        return 1
+    fi
+
+    PYTHON_BIN="$(command -v "$py_cmd")"
+    ok "Python: $("$py_cmd" --version 2>&1) ($PYTHON_BIN)"
+    return 0
+}
+
+check_cmake() {
+    if ! command -v cmake &>/dev/null; then
+        fail "cmake 未找到（编译 nvbandwidth 需要 >= 3.18）"
+        echo "  → apt install cmake"
+        return 1
+    fi
+    local cmake_ver
+    cmake_ver=$(cmake --version | head -1 | grep -oP '[0-9]+\.[0-9]+')
+    local cmake_major cmake_minor
+    cmake_major="${cmake_ver%%.*}"
+    cmake_minor="${cmake_ver#*.}"
+    if [[ "$cmake_major" -lt 3 ]] || { [[ "$cmake_major" -eq 3 ]] && [[ "$cmake_minor" -lt 18 ]]; }; then
+        fail "cmake $cmake_ver < 3.18（nvbandwidth 需要 >= 3.18）"
+        echo "  → 升级 cmake: pip install cmake 或从源码安装"
+        return 1
+    fi
+    ok "cmake: $cmake_ver"
+    return 0
+}
+
+check_compiler() {
+    if ! command -v gcc &>/dev/null || ! command -v g++ &>/dev/null; then
+        fail "gcc/g++ 未找到"
+        echo "  → apt install build-essential"
+        return 1
+    fi
+    local gcc_ver
+    gcc_ver=$(gcc -dumpversion 2>/dev/null)
+    ok "gcc/g++: $gcc_ver"
+    return 0
+}
+
+check_mpi() {
+    if command -v mpirun &>/dev/null || command -v mpiexec &>/dev/null; then
+        HAS_MPI=1
+        ok "MPI: $(mpirun --version 2>&1 | head -1)"
+    else
+        HAS_MPI=0
+        warn "mpirun 未找到（nccl-tests 将不使用 MPI 模式）"
+        echo "  → apt install openmpi-bin libopenmpi-dev"
+    fi
+    return 0
+}
+
+check_nccl_dev() {
+    if ldconfig -p 2>/dev/null | grep -q libnccl; then
+        HAS_NCCL_DEV=1
+        ok "libnccl: 已找到 (via ldconfig)"
+        return 0
+    fi
+    if [[ -f /usr/include/nccl.h ]] || dpkg -l libnccl-dev &>/dev/null 2>&1; then
+        HAS_NCCL_DEV=1
+        ok "libnccl-dev: 已安装"
+        return 0
+    fi
+    HAS_NCCL_DEV=0
+    warn "libnccl-dev 未找到（将跳过 nccl-tests 编译）"
+    echo "  → apt install libnccl-dev libnccl2"
+    return 0
 }
 
 install_system_deps() {
-    log "Installing system dependencies..."
+    log "安装系统依赖包..."
     if command -v apt-get &>/dev/null; then
         apt-get update -qq
         apt-get install -y -qq build-essential git cmake wget curl \
             openmpi-bin libopenmpi-dev openssh-client \
             infiniband-diags ibverbs-utils perftest \
             python3 python3-pip python3-venv \
-            2>/dev/null || warn "Some apt packages failed (may already be installed)"
-    elif command -v yum &>/dev/null; then
-        yum groupinstall -y "Development Tools" 2>/dev/null || true
-        yum install -y git cmake wget curl \
-            openmpi openmpi-devel openssh-clients \
-            infiniband-diags libibverbs-utils perftest \
-            python3 python3-pip \
-            2>/dev/null || warn "Some yum packages failed"
+            libnccl-dev libnccl2 \
+            2>/dev/null || warn "部分包安装失败（可能已安装）"
     elif command -v dnf &>/dev/null; then
         dnf groupinstall -y "Development Tools" 2>/dev/null || true
         dnf install -y git cmake wget curl \
             openmpi openmpi-devel openssh-clients \
             infiniband-diags libibverbs-utils perftest \
             python3 python3-pip \
-            2>/dev/null || warn "Some dnf packages failed"
+            2>/dev/null || warn "部分包安装失败"
+    elif command -v yum &>/dev/null; then
+        yum groupinstall -y "Development Tools" 2>/dev/null || true
+        yum install -y git cmake wget curl \
+            openmpi openmpi-devel openssh-clients \
+            infiniband-diags libibverbs-utils perftest \
+            python3 python3-pip \
+            2>/dev/null || warn "部分包安装失败"
     else
-        warn "Unsupported package manager. Install deps manually."
+        warn "未识别的包管理器，请手动安装依赖"
     fi
-    ok "System dependencies"
+    ok "系统依赖安装完成"
 }
 
-install_python_deps() {
-    log "Installing Python dependencies..."
-    pip3 install --quiet rich pyyaml 2>/dev/null || pip install --quiet rich pyyaml
-    ok "Python dependencies (rich, pyyaml)"
+validate_environment() {
+    banner "阶段 1/6: 环境校验"
+
+    local errors=0
+
+    check_nvidia_smi    || ((errors++))
+    detect_gpu_and_driver || true
+    detect_cuda_version || ((errors++))
+    check_python        || ((errors++))
+    check_cmake         || ((errors++))
+    check_compiler      || ((errors++))
+    check_mpi           || true
+    check_nccl_dev      || true
+
+    echo ""
+    if [[ $errors -gt 0 ]]; then
+        fail "环境校验发现 $errors 个必要组件缺失"
+        if [[ $FLAG_INSTALL_SYS_DEPS -eq 1 ]]; then
+            log "检测到 --install-system-deps，尝试安装..."
+            install_system_deps
+            # 重新校验
+            errors=0
+            check_python  || ((errors++))
+            check_cmake   || ((errors++))
+            check_compiler || ((errors++))
+            check_mpi     || true
+            check_nccl_dev || true
+            if [[ $errors -gt 0 ]]; then
+                die "安装系统包后仍有 $errors 个组件缺失，请手动解决"
+            fi
+        else
+            echo ""
+            echo "  提示: 加 --install-system-deps 参数可自动安装缺失的系统包"
+            echo "  或手动运行上面提示的 apt install 命令后重试"
+            die "环境校验未通过"
+        fi
+    fi
+    ok "环境校验通过"
 }
 
-install_nvbandwidth() {
-    log "Installing nvbandwidth..."
+# ─── 阶段 2: 安装 uv ──────────────────────────────────────────────────────────
+ensure_uv() {
+    banner "阶段 2/6: 确保 uv 可用"
+
+    # 检查已有的 uv
+    if command -v uv &>/dev/null; then
+        ok "uv 已安装: $(uv --version 2>&1)"
+        return 0
+    fi
+
+    # 检查常见位置
+    for p in "$HOME/.local/bin/uv" "$HOME/.cargo/bin/uv" /usr/local/bin/uv; do
+        if [[ -x "$p" ]]; then
+            export PATH="$(dirname "$p"):$PATH"
+            ok "uv 已找到: $p"
+            return 0
+        fi
+    done
+
+    log "正在安装 uv..."
+    if ! curl -LsSf https://astral.sh/uv/install.sh | sh 2>/dev/null; then
+        die "uv 安装失败。请手动安装: https://docs.astral.sh/uv/getting-started/installation/"
+    fi
+
+    # 将 uv 加入 PATH
+    export PATH="$HOME/.local/bin:$HOME/.cargo/bin:$PATH"
+
+    if ! command -v uv &>/dev/null; then
+        die "uv 安装后仍无法找到。请检查 PATH。"
+    fi
+    ok "uv 安装成功: $(uv --version 2>&1)"
+}
+
+# ─── 阶段 3: Python 虚拟环境 ──────────────────────────────────────────────────
+setup_python_venv() {
+    banner "阶段 3/6: 创建 Python 虚拟环境"
+
+    local venv_dir="$INSTALL_DIR/.venv"
+
+    # 检查已有 venv
+    if [[ -x "$venv_dir/bin/python" ]] && [[ $FLAG_REBUILD -eq 0 ]]; then
+        local existing_ver
+        existing_ver=$("$venv_dir/bin/python" -c "import sys; print(f'{sys.version_info.major}.{sys.version_info.minor}')" 2>/dev/null || echo "0.0")
+        local ev_major="${existing_ver%%.*}"
+        local ev_minor="${existing_ver#*.}"
+        if [[ "$ev_major" -ge 3 ]] && [[ "$ev_minor" -ge 10 ]]; then
+            ok "虚拟环境已存在: $venv_dir (Python $existing_ver)"
+        else
+            log "已有 venv 的 Python 版本过低 ($existing_ver)，重建中..."
+            rm -rf "$venv_dir"
+        fi
+    fi
+
+    # 创建 venv
+    if [[ ! -x "$venv_dir/bin/python" ]]; then
+        log "创建虚拟环境: $venv_dir"
+        uv venv "$venv_dir" --python "$PYTHON_BIN"
+        ok "虚拟环境创建成功"
+    fi
+
+    # 安装项目依赖
+    log "安装 Python 依赖（rich、pyyaml、numpy）..."
+    uv pip install --python "$venv_dir/bin/python" \
+        -e "$PROJECT_DIR" 2>&1 | { [[ $VERBOSE -eq 1 ]] && cat || tail -1; } || true
+    ok "项目依赖安装完成"
+
+    # 安装 PyTorch
+    if [[ $FLAG_SKIP_PYTORCH -eq 1 ]]; then
+        warn "跳过 PyTorch 安装（--skip-pytorch）"
+    else
+        # 检查是否已有 torch
+        if "$venv_dir/bin/python" -c "import torch" &>/dev/null && [[ $FLAG_REBUILD -eq 0 ]]; then
+            local torch_ver
+            torch_ver=$("$venv_dir/bin/python" -c "import torch; print(torch.__version__)" 2>/dev/null)
+            ok "PyTorch 已安装: $torch_ver"
+        else
+            local index_url="https://download.pytorch.org/whl/${CUDA_TAG}"
+            log "安装 PyTorch (CUDA $CUDA_TAG): $index_url"
+            log "（下载较大，请耐心等待...）"
+            uv pip install --python "$venv_dir/bin/python" \
+                "torch>=2.1.0" --index-url "$index_url" \
+                2>&1 | { [[ $VERBOSE -eq 1 ]] && cat || tail -3; } || {
+                    warn "PyTorch 安装失败，可稍后手动安装:"
+                    echo "  source $INSTALL_DIR/env.sh"
+                    echo "  uv pip install torch --index-url $index_url"
+                }
+            if "$venv_dir/bin/python" -c "import torch" &>/dev/null; then
+                local torch_ver
+                torch_ver=$("$venv_dir/bin/python" -c "import torch; print(torch.__version__)" 2>/dev/null)
+                ok "PyTorch 安装成功: $torch_ver"
+            fi
+        fi
+    fi
+}
+
+# ─── 阶段 4: 编译原生工具 ─────────────────────────────────────────────────────
+build_nvbandwidth() {
     local src="$INSTALL_DIR/nvbandwidth"
-    if [[ -x "$src/nvbandwidth" ]]; then
-        ok "nvbandwidth already installed at $src/nvbandwidth"
-        return
+
+    # 幂等检查
+    if [[ -x "$src/nvbandwidth" ]] && [[ $FLAG_REBUILD -eq 0 ]]; then
+        ok "nvbandwidth: 已编译 ($src/nvbandwidth)"
+        return 0
     fi
 
-    mkdir -p "$INSTALL_DIR"
-    git clone --depth 1 https://github.com/NVIDIA/nvbandwidth.git "$src" 2>/dev/null
-    cd "$src"
-    mkdir -p build && cd build
-    cmake .. -DCMAKE_BUILD_TYPE=Release 2>/dev/null
-    make -j"$JOBS" 2>/dev/null
-    if [[ -x "$src/build/nvbandwidth" ]]; then
-        cp "$src/build/nvbandwidth" "$src/nvbandwidth"
-        ok "nvbandwidth installed at $src/nvbandwidth"
+    log "编译 nvbandwidth..."
+    (
+        set -e
+        # 清理 / 克隆
+        if [[ $FLAG_REBUILD -eq 1 ]] && [[ -d "$src" ]]; then
+            rm -rf "$src"
+        fi
+        if [[ -d "$src/.git" ]]; then
+            cd "$src" && git pull --ff-only 2>/dev/null || true
+        elif [[ -d "$src" ]]; then
+            rm -rf "$src"
+            git clone --depth 1 https://github.com/NVIDIA/nvbandwidth.git "$src"
+        else
+            git clone --depth 1 https://github.com/NVIDIA/nvbandwidth.git "$src"
+        fi
+
+        cd "$src"
+        mkdir -p build && cd build
+        cmake .. -DCMAKE_BUILD_TYPE=Release 2>&1 | { [[ $VERBOSE -eq 1 ]] && cat || tail -3; }
+        make -j"$JOBS" 2>&1 | { [[ $VERBOSE -eq 1 ]] && cat || tail -3; }
+
+        if [[ -x ./nvbandwidth ]]; then
+            cp ./nvbandwidth "$src/nvbandwidth"
+        fi
+    )
+
+    if [[ -x "$src/nvbandwidth" ]]; then
+        ok "nvbandwidth: 编译成功"
     else
-        warn "nvbandwidth build failed. Try building manually in $src"
+        warn "nvbandwidth: 编译失败（非致命，可手动编译）"
     fi
 }
 
-install_nccl_tests() {
-    log "Installing nccl-tests..."
+build_nccl_tests() {
     local src="$INSTALL_DIR/nccl-tests"
-    if [[ -x "$src/build/all_reduce_perf" ]]; then
-        ok "nccl-tests already installed at $src/build/"
-        return
+
+    if [[ -x "$src/build/all_reduce_perf" ]] && [[ $FLAG_REBUILD -eq 0 ]]; then
+        ok "nccl-tests: 已编译 ($src/build/)"
+        return 0
     fi
 
-    mkdir -p "$INSTALL_DIR"
-    git clone --depth 1 https://github.com/NVIDIA/nccl-tests.git "$src" 2>/dev/null
-    cd "$src"
+    if [[ $HAS_NCCL_DEV -eq 0 ]]; then
+        warn "nccl-tests: 跳过（libnccl-dev 未安装）"
+        return 0
+    fi
+
+    local cuda_home="${CUDA_HOME:-/usr/local/cuda}"
+    if [[ ! -d "$cuda_home/include" ]]; then
+        warn "nccl-tests: 跳过（CUDA_HOME=$cuda_home 无效）"
+        return 0
+    fi
+
+    log "编译 nccl-tests..."
+    (
+        set -e
+        if [[ $FLAG_REBUILD -eq 1 ]] && [[ -d "$src" ]]; then
+            rm -rf "$src"
+        fi
+        if [[ -d "$src/.git" ]]; then
+            cd "$src" && git pull --ff-only 2>/dev/null || true
+        elif [[ -d "$src" ]]; then
+            rm -rf "$src"
+            git clone --depth 1 https://github.com/NVIDIA/nccl-tests.git "$src"
+        else
+            git clone --depth 1 https://github.com/NVIDIA/nccl-tests.git "$src"
+        fi
+
+        cd "$src"
+        if [[ $HAS_MPI -eq 1 ]]; then
+            make MPI=1 MPI_HOME=/usr CUDA_HOME="$cuda_home" -j"$JOBS" \
+                2>&1 | { [[ $VERBOSE -eq 1 ]] && cat || tail -3; }
+        else
+            make CUDA_HOME="$cuda_home" -j"$JOBS" \
+                2>&1 | { [[ $VERBOSE -eq 1 ]] && cat || tail -3; }
+        fi
+    )
+
+    if [[ -x "$src/build/all_reduce_perf" ]]; then
+        ok "nccl-tests: 编译成功"
+    else
+        warn "nccl-tests: 编译失败（非致命）"
+    fi
+}
+
+build_gpu_burn() {
+    local src="$INSTALL_DIR/gpu-burn"
+
+    if [[ -x "$src/gpu_burn" ]] && [[ $FLAG_REBUILD -eq 0 ]]; then
+        ok "gpu-burn: 已编译 ($src/gpu_burn)"
+        return 0
+    fi
 
     local cuda_home="${CUDA_HOME:-/usr/local/cuda}"
     if [[ ! -d "$cuda_home" ]]; then
-        warn "CUDA_HOME not found at $cuda_home. Set CUDA_HOME env var."
-        return
+        warn "gpu-burn: 跳过（CUDA_HOME=$cuda_home 不存在）"
+        return 0
     fi
 
-    make MPI=1 MPI_HOME=/usr -j"$JOBS" 2>/dev/null || \
-        make CUDA_HOME="$cuda_home" -j"$JOBS" 2>/dev/null || \
-        warn "nccl-tests build failed. Try: cd $src && make MPI=1"
-
-    if [[ -x "$src/build/all_reduce_perf" ]]; then
-        ok "nccl-tests installed at $src/build/"
-    else
-        warn "nccl-tests build incomplete"
-    fi
-}
-
-install_gpu_burn() {
-    log "Installing gpu-burn..."
-    local src="$INSTALL_DIR/gpu-burn"
-    if [[ -x "$src/gpu_burn" ]]; then
-        ok "gpu-burn already installed at $src/gpu_burn"
-        return
-    fi
-
-    mkdir -p "$INSTALL_DIR"
-    git clone --depth 1 https://github.com/wilicc/gpu-burn.git "$src" 2>/dev/null
-    cd "$src"
-    make -j"$JOBS" 2>/dev/null || warn "gpu-burn build failed"
-    if [[ -x "$src/gpu_burn" ]]; then
-        ok "gpu-burn installed at $src/gpu_burn"
-    else
-        warn "gpu-burn build incomplete"
-    fi
-}
-
-check_dcgm() {
-    log "Checking DCGM..."
-    if command -v nv-hostengine &>/dev/null || command -v dcgmi &>/dev/null; then
-        ok "DCGM already installed"
-        return
-    fi
-    if dpkg -l datacenter-gpu-manager &>/dev/null 2>&1; then
-        ok "DCGM package installed"
-        return
-    fi
-    warn "DCGM not found. Install from: https://docs.nvidia.com/datacenter/dcgm/latest/installation-guide.html"
-    warn "  Ubuntu: sudo apt install datacenter-gpu-manager"
-    warn "  Or:     curl -fsSL https://deb.nvidia.com/datacenter-gpu-manager/gpgkey | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-dcgm.gpg"
-}
-
-check_rdma_tools() {
-    log "Checking RDMA tools..."
-    local found=0
-    for tool in ib_write_bw ib_read_bw ib_write_lat ib_read_lat ibstat ibv_devinfo; do
-        if command -v "$tool" &>/dev/null; then
-            found=$((found + 1))
-        else
-            warn "  $tool not found (install: perftest infiniband-diags)"
+    log "编译 gpu-burn..."
+    (
+        set -e
+        if [[ $FLAG_REBUILD -eq 1 ]] && [[ -d "$src" ]]; then
+            rm -rf "$src"
         fi
-    done
-    if [[ $found -gt 0 ]]; then
-        ok "$found/$RDMA_TOOL_COUNT RDMA tools found" 2>/dev/null || ok "Some RDMA tools found"
+        if [[ -d "$src/.git" ]]; then
+            cd "$src" && git pull --ff-only 2>/dev/null || true
+        elif [[ -d "$src" ]]; then
+            rm -rf "$src"
+            git clone --depth 1 https://github.com/wilicc/gpu-burn.git "$src"
+        else
+            git clone --depth 1 https://github.com/wilicc/gpu-burn.git "$src"
+        fi
+
+        cd "$src"
+        make CUDA_PATH="$cuda_home" -j"$JOBS" \
+            2>&1 | { [[ $VERBOSE -eq 1 ]] && cat || tail -3; }
+    )
+
+    if [[ -x "$src/gpu_burn" ]]; then
+        ok "gpu-burn: 编译成功"
+    else
+        warn "gpu-burn: 编译失败（非致命）"
     fi
 }
 
-print_summary() {
-    echo ""
-    echo "=========================================="
-    echo " GPU Test Suite - Installation Summary"
-    echo "=========================================="
-    echo ""
-    echo " Install directory: $INSTALL_DIR"
-    echo ""
-    echo " Tools status:"
+build_native_tools() {
+    banner "阶段 4/6: 编译原生工具"
+    build_nvbandwidth
+    build_nccl_tests
+    build_gpu_burn
+}
 
-    for tool_path in \
+# ─── 阶段 5: 生成激活脚本 ─────────────────────────────────────────────────────
+generate_env_sh() {
+    banner "阶段 5/6: 生成环境脚本"
+
+    local env_file="$INSTALL_DIR/env.sh"
+    cat > "$env_file" << 'ENVEOF'
+#!/usr/bin/env bash
+# GPU Test Suite 环境激活脚本
+# 用法: source /opt/gpu-test-tools/env.sh
+
+export GPU_TOOLS_DIR="__INSTALL_DIR__"
+export CUDA_HOME="${CUDA_HOME:-/usr/local/cuda}"
+
+# 激活 Python 虚拟环境
+if [[ -f "$GPU_TOOLS_DIR/.venv/bin/activate" ]]; then
+    source "$GPU_TOOLS_DIR/.venv/bin/activate"
+fi
+
+# 编译工具加入 PATH
+export PATH="$GPU_TOOLS_DIR/nvbandwidth:$PATH"
+export PATH="$GPU_TOOLS_DIR/nccl-tests/build:$PATH"
+export PATH="$GPU_TOOLS_DIR/gpu-burn:$PATH"
+export PATH="$CUDA_HOME/bin:$PATH"
+
+# 库路径
+export LD_LIBRARY_PATH="$CUDA_HOME/lib64:${LD_LIBRARY_PATH:-}"
+ENVEOF
+
+    # 替换占位符
+    sed -i "s|__INSTALL_DIR__|$INSTALL_DIR|g" "$env_file"
+    chmod +x "$env_file"
+    ok "env.sh 已生成: $env_file"
+
+    # 生成 run-gpu-tests 运行器
+    local wrapper="$INSTALL_DIR/run-gpu-tests"
+    cat > "$wrapper" << WRAPEOF
+#!/usr/bin/env bash
+# GPU Test Suite 一键运行器
+# 用法: /opt/gpu-test-tools/run-gpu-tests --test all
+
+SCRIPT_DIR="\$(cd "\$(dirname "\${BASH_SOURCE[0]}")" && pwd)"
+source "\$SCRIPT_DIR/env.sh"
+exec python3 "$PROJECT_DIR/gpu_tester.py" "\$@"
+WRAPEOF
+    chmod +x "$wrapper"
+    ok "run-gpu-tests 已生成: $wrapper"
+}
+
+# ─── 阶段 6: 打印总结 ─────────────────────────────────────────────────────────
+print_summary() {
+    banner "阶段 6/6: 安装总结"
+
+    echo -e "${BOLD}安装目录:${NC} $INSTALL_DIR"
+    echo ""
+    echo -e "${BOLD}组件状态:${NC}"
+
+    # Python 虚拟环境
+    if [[ -x "$INSTALL_DIR/.venv/bin/python" ]]; then
+        local py_ver
+        py_ver=$("$INSTALL_DIR/.venv/bin/python" --version 2>&1)
+        echo -e "  ${GREEN}✓${NC} Python venv: $py_ver"
+    else
+        echo -e "  ${RED}✗${NC} Python venv: 未创建"
+    fi
+
+    # PyTorch
+    if "$INSTALL_DIR/.venv/bin/python" -c "import torch" &>/dev/null 2>&1; then
+        local tv
+        tv=$("$INSTALL_DIR/.venv/bin/python" -c "import torch; print(f'{torch.__version__} (CUDA {torch.version.cuda})')" 2>/dev/null)
+        echo -e "  ${GREEN}✓${NC} PyTorch: $tv"
+    else
+        echo -e "  ${YELLOW}○${NC} PyTorch: 未安装"
+    fi
+
+    # 编译工具
+    for tool_info in \
         "$INSTALL_DIR/nvbandwidth/nvbandwidth:nvbandwidth" \
         "$INSTALL_DIR/nccl-tests/build/all_reduce_perf:nccl-tests" \
         "$INSTALL_DIR/gpu-burn/gpu_burn:gpu-burn"; do
-        path="${tool_path%%:*}"
-        name="${tool_path##*:}"
+        local path="${tool_info%%:*}"
+        local name="${tool_info##*:}"
         if [[ -x "$path" ]]; then
-            echo -e "  [${GREEN}✓${NC}] $name"
+            echo -e "  ${GREEN}✓${NC} $name"
         else
-            echo -e "  [${YELLOW}?${NC}] $name (not built)"
+            echo -e "  ${YELLOW}○${NC} $name (未编译)"
         fi
     done
 
-    echo ""
-    echo " System tools:"
-    for cmd in nvidia-smi mpirun nvbandwidth ib_write_bw dcgmi; do
-        if command -v "$cmd" &>/dev/null; then
-            echo -e "  [${GREEN}✓${NC}] $cmd"
-        else
-            echo -e "  [${YELLOW}-${NC}] $cmd (not found)"
+    # RDMA 工具（系统级）
+    local rdma_found=0
+    for tool in ib_write_bw ib_read_bw ibstat; do
+        if command -v "$tool" &>/dev/null; then
+            ((rdma_found++))
         fi
     done
+    if [[ $rdma_found -gt 0 ]]; then
+        echo -e "  ${GREEN}✓${NC} RDMA 工具: $rdma_found/3 可用"
+    else
+        echo -e "  ${YELLOW}○${NC} RDMA 工具: 未安装 (apt install perftest infiniband-diags)"
+    fi
 
     echo ""
-    echo " Usage:"
-    echo "   python3 gpu_tester.py              # Interactive menu"
-    echo "   python3 gpu_tester.py --test all    # Full suite"
+    echo -e "${BOLD}使用方法:${NC}"
+    echo ""
+    echo "  # 方式一: source 激活后使用"
+    echo "  source $INSTALL_DIR/env.sh"
+    echo "  python3 $PROJECT_DIR/gpu_tester.py --test all"
+    echo ""
+    echo "  # 方式二: 一键运行"
+    echo "  $INSTALL_DIR/run-gpu-tests --test all"
+    echo "  $INSTALL_DIR/run-gpu-tests --test health"
+    echo "  $INSTALL_DIR/run-gpu-tests              # 交互式菜单"
     echo ""
 }
 
+# ─── 主函数 ───────────────────────────────────────────────────────────────────
 main() {
+    parse_args "$@"
+
     echo ""
-    echo "=========================================="
-    echo " GPU Test Suite - Dependency Installer"
-    echo "=========================================="
+    echo -e "${BOLD}${CYAN}╔══════════════════════════════════════════════════╗${NC}"
+    echo -e "${BOLD}${CYAN}║   GPU Test Suite — 一键安装                      ║${NC}"
+    echo -e "${BOLD}${CYAN}║   环境隔离 · 自动检测 · 完整部署                ║${NC}"
+    echo -e "${BOLD}${CYAN}╚══════════════════════════════════════════════════╝${NC}"
+    echo ""
+    log "安装目录: $INSTALL_DIR"
+    log "项目目录: $PROJECT_DIR"
     echo ""
 
-    check_root
-    detect_gpu
-
-    mkdir -p "$INSTALL_DIR"
-
-    install_system_deps
-    install_python_deps
-    install_nvbandwidth
-    install_nccl_tests
-    install_gpu_burn
-    check_dcgm
-    check_rdma_tools
-
+    check_permissions
+    validate_environment
+    ensure_uv
+    setup_python_venv
+    build_native_tools
+    generate_env_sh
     print_summary
+
+    echo -e "${GREEN}${BOLD}安装完成！${NC}"
 }
 
 main "$@"
diff --git a/requirements.txt b/requirements.txt
index 284cb79..4be4c43 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,2 +1,3 @@
 rich>=13.0
 pyyaml>=6.0
+numpy>=1.24