以下流程基于 WSL 2 + Ubuntu 24.04，兼顾未来 RTX 50 系列（预估 SM 9.x）、现有 H100 （SM 9.0）与 A100 （SM 8.0）等多代 GPU，同时满足 Python 与 C++ 混合开发（PyTorch + LibTorch）。关键要点是：选用 ≥ CUDA 12.8/12.9 的最新正式版，它向下兼容 SM 8.0，首次完整支持 SM 9.x；Windows 侧仅装一次支持 WSL 的 NVIDIA 驱动；Linux 侧只装 toolkit 与 cuDNN（不要装驱动）；PyTorch/LibTorch 版本必须与 CUDA 小版本对齐。完成后即可在同一环境中无缝切换 RTX 50-series、H100、A100 等卡。developer.nvidia.com developer.nvidia.com developer.nvidia.com

1 环境与版本策略

计算能力及最低 CUDA 版本见官方表格 developer.nvidia.com developer.nvidia.com forums.developer.nvidia.com。因此，统一安装 CUDA 12.9（或之后发布的 13.x 稳定版）即可让上述卡全部工作，并为后续 RTX 50-series 留足余量。

2 Windows 11/10 主机：安装支持 WSL GPU 的驱动

powershell
# 管理员 PowerShell
wsl --install                 # 如已装 WSL 可跳过
winget install --id=NVIDIA.NVIDIAApp -e

# 在 NVIDIA App 中选择最新正式版 (≥ v560+) 驱动

• 官方 WSL 驱动与普通 GeForce/RTX Enterprise 驱动已合并，无需单独“预览”包。developer.nvidia.com learn.microsoft.com
• 安装完成后重启；nvidia-smi 在 Windows PowerShell 中应能看到 GPU。

3 Ubuntu 24.04(WSL) 内部：安装 CUDA 12.9 Toolkit

bash
# 3.1 移除旧 key（若曾用过）
sudo apt-key del 7fa2af80 || true

# 3.2 添加官方仓库 pin
wget https://developer.download.nvidia.com/compute/cuda/repos/wsl-ubuntu/x86_64/cuda-wsl-ubuntu.pin
sudo mv cuda-wsl-ubuntu.pin /etc/apt/preferences.d/cuda-repository-pin-600

# 3.3 安装 keyring 与 Toolkit
sudo apt-get update
sudo apt-get install -y cuda-toolkit-12-9

• 不要在 WSL 里安装 cuda-drivers 包，否则会错误尝试加载 Linux 驱动。� citedocs.nvidia.com docs.nvidia.com docs.nvidia.com
• Toolkit 安装路径默认 /usr/local/cuda-12.9，并自动创建 /usr/local/cuda 指向当前活动版本。docs.nvidia.com

可通过 sudo update-alternatives --config cuda 在多版本 Toolkit 间切换。docs.nvidia.com

4 安装 cuDNN 9.x（与 CUDA 小版本一致）

bash
# 网络安装方式（官方 .deb）
sudo apt-get install -y libcudnn9-dev=9.1.*+cuda12.9

• Ubuntu 24.04 已提供独立 deb 包；安装时会自动拉取匹配的 libcudnn9、libcudnn9-samples。developer.nvidia.com documentation.ubuntu.com
• 若需多版本并存，可使用 libcudnn*-dev-*-cuda12-8 与 12-9 并行。

5 Python 端：安装与 CUDA 12.9 匹配的 PyTorch

bash
# 建议在 virtualenv / conda 中执行
pip install torch==2.7.0+cu129 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu129

• 版本选择器见官方页面；确保 “Compute Platform” 选 CUDA 12.9。pytorch.org pytorch.org
• 验证：

python
import torch, torchvision
print(torch.__version__, torch.cuda.is_available(), torch.cuda.get_device_properties(0).major)

6 C++ 端：配置 LibTorch 并与 CMake 集成

1. 下载 libtorch-cxx11-abi-shared-with-deps-2.7.0+cu129.zip。pytorch.org
2. 解压到 /opt/libtorch-cu129，并设置环境

bash
export Torch_DIR=/opt/libtorch-cu129
export LD_LIBRARY_PATH=$Torch_DIR/lib:$LD_LIBRARY_PATH

3. CMakeLists.txt 示例

cmake
cmake_minimum_required(VERSION 3.20)
project(myapp LANGUAGES CXX)
find_package(Torch REQUIRED PATHS $ENV{Torch_DIR})
add_executable(myapp main.cpp)
target_link_libraries(myapp "${TORCH_LIBRARIES}")
target_compile_options(myapp PRIVATE
  -std=c++17
  -gencode=arch=compute_80,code=sm_80      # A100
  -gencode=arch=compute_90,code=sm_90      # H100 / RTX 50
)
set_property(TARGET myapp PROPERTY CXX_STANDARD 17)

示例及常见问题可参考官方 C++ 教程与社区帖。docs.pytorch.org discuss.pytorch.org docs.pytorch.org

7 跨架构兼容要点（A100 / H100 / RTX 50）

• 只要 驱动 ≥ R560、CUDA ≥ 11.8，同时支持 SM 8.0–9.x 二进制。forums.developer.nvidia.com developer.nvidia.com developer.nvidia.com
• 数据中心卡 H100/A100 可通过 MIG 在容器内分区；Docker 请加入 --gpus all 与 --cap-add=sys_nice 等参数。developer.nvidia.com

8 多版本管理与容器化

• update-alternatives：管理 /usr/local/cuda 软链。docs.nvidia.com
• nvidia-container-toolkit：在 WSL2 Ubuntu 中直接 sudo apt-get install nvidia-docker2，然后用官方 nvcr.io/nvidia/pytorch:24.05-py3 容器即可。docs.nvidia.com
• 对需要固定驱动版本的生产环境，推荐 WSLg + Docker Desktop GPU pass-through，可确保 Windows 与 Linux 层驱动一致。learn.microsoft.com

9 常见问题排查

10 快速自测

bash
# Linux
nvcc --version         # 应显示 release 12.9
nvidia-smi             # Driver ≥ 560.* CUDA Version 12.9
/usr/local/cuda/samples/1_Utilities/deviceQuery/deviceQuery

成功输出 Result = PASS 及 SM 90/80/89 等信息，即表示整套链路就绪。至此，您已拥有一套同时驱动 RTX 50-series / H100 / A100 的 CUDA + cuDNN + (Py)Torch +C++ 开发环境，可直接开始混合编程与模型部署。祝开发顺利!

本质上，你只需要 3 步 就能让这份 CMakeLists.txt 同时 “看见” LibTorch 和 cuDNN：① 让 CMake 找到它们（find_package）；② 用官方导出的 INTERFACE target（Torch::Torch、CuDNN::CuDNN）来编译/链接；③ 确保运行时能搜到它们（RPATH 或 LD_LIBRARY_PATH）。下面先给出一段 最小可行 diff，再逐段解释要点和常见陷阱。

diff
@@
-# --- CUDA Architecture ---
-set(CMAKE_CUDA_ARCHITECTURES "120")
+# --- CUDA Architecture ---
+# 90=H100/RTX-50，80=A100，89=RTX-40。如只在本机编译可用 native 自动探测。
+set(CMAKE_CUDA_ARCHITECTURES 90;80;89)

-# --- Ensure CUDA Toolkit targets (nvToolsExt, etc.) are available ---
-find_package(CUDAToolkit REQUIRED)
+# --- Ensure CUDA Toolkit targets ---
+find_package(CUDAToolkit REQUIRED)

-# --- LibTorch Configuration ---
+# --- LibTorch ---
+# 建议：提前 export Torch_DIR=/opt/libtorch-cu129/share/cmake/Torch
+find_package(Torch REQUIRED)        # 生成 Torch::Torch 等 INTERFACE target
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${TORCH_CXX_FLAGS}")

+# --- cuDNN ---
+# Debian/Ubuntu 官方 .deb 会安装 FindCUDNN.cmake；若自行解压版可手动给根目录：
+#   cmake -DCUDNN_ROOT=/usr/local/cuda .
+find_package(CUDNN REQUIRED)        # 生成 CuDNN::CuDNN 或 ${CUDNN_LIBRARIES}

@@
-target_link_libraries(qm_core PUBLIC
-    ${TORCH_LIBRARIES}
-    ${CUDNN_LIBRARIES}
-)
+target_link_libraries(qm_core PUBLIC
+    Torch::Torch           # 自动带上依赖 (cudart,c10,torch_cpu,torch_cuda…)
+    CuDNN::CuDNN           # cuDNN 9.x 多子库变体一次性全部解决
+)

@@
-set_target_properties(test_cudnn PROPERTIES
-    BUILD_RPATH "${LIBTORCH_INSTALL_PREFIX}/lib"
-)
+# 让可执行文件运行时能找到 libtorch/cuDNN
+set_target_properties(test_cudnn PROPERTIES
+    BUILD_RPATH "$ORIGIN;$ORIGIN/../lib"  # 或改用 INSTALL_RPATH
+)

1. 让 CMake 找到 LibTorch

• LibTorch 不是用 find_library，而是配置包，官方生成 TorchConfig.cmake；只要设置好 Torch_DIR 或 CMAKE_PREFIX_PATH 即可，让 find_package(Torch REQUIRED) 自动导入 Torch::Torch, ${TORCH_CXX_FLAGS} 等变量。docs.pytorch.org stackoverflow.com
• Torch_DIR 路径可用 Python 查询：

  bash
  python - <<'PY'
  import torch, pathlib, sys
  print(pathlib.Path(torch.utils.cmake_prefix_path))
  PY
  ``` 这会输出 `.../libtorch/share/cmake/Torch`。[stackoverflow.com](https://stackoverflow.com/questions/76443205/what-does-the-cmake-variable-torch-dir-do-how-do-i-find-out-what-directory-to-h?utm_source=chatgpt.com)  

• 将该路径写进 CMAKE_PREFIX_PATH 或在 shell 中 export Torch_DIR=...，再执行 cmake ..。discourse.cmake.org

2. 配置 cuDNN

• 新 cuDNN (9.x) 在 /usr/include + /usr/lib/x86_64-linux-gnu 安装时会带 FindCUDNN.cmake；直接 find_package(CUDNN REQUIRED) 即可生成 CuDNN::CuDNN 和 ${CUDNN_LIBRARIES} 变量。vcpkg.link docs.nvidia.com
• 若手动解压版，把根目录通过 -DCUDNN_ROOT=/path/to/cudnn 指给模块或在环境变量里写 CUDNN_ROOT.stackoverflow.com docs.nvidia.com
• cuDNN 9 拆分了若干子库（libcudnn_ops*, libcudnn_graph 等），用官方 CMake 脚本比手动 find_library 全得多，避免遗漏。docs.nvidia.com

3. 修正 CUDA 架构设置

• "120" 指 Fermi（计算能力 1.2），早已淘汰；H100/RTX-50 是 sm_90，A100 是 sm_80，RTX Ada 是 sm_89。cmake.org
• 你可以写 set(CMAKE_CUDA_ARCHITECTURES native) 让 CMake 3.24+ 自行探测，或显式列出多档；逗号/分号皆可。stackoverflow.com

4. 链接/编译细节

4.1 INTERFACE target 比变量安全

Torch::Torch 会自动传播需要的 -D_GLIBCXX_USE_CXX11_ABI, -fPIC, 以及全部依赖库；不必再把 ${TORCH_INCLUDE_DIRS} / ${TORCH_LIBRARIES} 散落各处。stackoverflow.com discuss.pytorch.org

4.2 RPATH / 运行时搜索

把 build-rpath 设为 $ORIGIN（可执行文件目录）或 $ORIGIN/../lib，就不用每次 export LD_LIBRARY_PATH。这是 CMake 推荐做法，用 BUILD_RPATH 或 INSTALL_RPATH 属性完成。cmake.org duerrenberger.dev

5. 整体思路小结

1. 安装：/opt/libtorch-cu129 + sudo apt install libcudnn9-dev=9.1.*+cuda12.9。
2. 环境：

   bash
   export Torch_DIR=/opt/libtorch-cu129/share/cmake/Torch
   export CUDNN_ROOT=/usr

3. CMake：上述 diff 中的改动即可；其余保持原有项目结构不变。
4. cmake -B build -DCMAKE_BUILD_TYPE=Release ➜ cmake --build build -j ➜ ctest --test-dir build；test_cudnn 应报告 CUDA = true, cuDNN version = 9300（示例）。

这样就完成了 WSL Ubuntu 24.04 环境下 LibTorch + cuDNN 的编译与运行，并兼容 A100/H100/RTX-50 系列。祝编译顺利!

Below is a self-contained, ready-to-build CMakeLists.txt that activates LibTorch and cuDNN in your WSL 2 / Ubuntu 24.04 environment.
All comments are in English and I have hard-wired your Torch_DIR to

/home/wsmxcz/miniconda3/lib/python3.13/site-packages/torch/share/cmake

so CMake can always locate LibTorch without extra environment variables.

cmake
###############################################################################
#  Top-level CMake configuration for the CheCI project
#  • C++20 core code                       (static library: qm_core)
#  • LibTorch (PyTorch C++ API)            (interface target: Torch::Torch)
#  • cuDNN 9.x                             (interface target: CuDNN::CuDNN)
#  • CUDA architectures for A100 / H100 / RTX-50 / RTX-40
#  • Optional OpenMP                       (finds system compiler flags)
###############################################################################

# ---------- Minimum CMake version and project ----------
cmake_minimum_required(VERSION 3.20)          # 3.20+ gives better CUDA support
project(CheCI LANGUAGES CXX CUDA)

# ---------- C++ standard ----------
set(CMAKE_CXX_STANDARD 20)
set(CMAKE_CXX_STANDARD_REQUIRED ON)
set(CMAKE_CXX_EXTENSIONS OFF)

# ---------- Compiler warnings ----------
if (CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang")
    add_compile_options(
        -Wall -Wextra -Wpedantic -Wshadow -Wnon-virtual-dtor
        -Wold-style-cast -Wcast-align -Wunused -Woverloaded-virtual
        -Wconversion -Wsign-conversion -Wnull-dereference
        -Wdouble-promotion -Wformat=2
    )
elseif (MSVC)
    add_compile_options(/W4 /EHsc)
endif()

# ---------- Build-type specific flags ----------
set(CMAKE_CXX_FLAGS_RELEASE   "${CMAKE_CXX_FLAGS_RELEASE} -O3 -DNDEBUG")
set(CMAKE_CXX_FLAGS_DEBUG     "${CMAKE_CXX_FLAGS_DEBUG}   -g  -O0")

# ---------- CUDA architectures ----------
# 90 = H100 / future RTX-50, 80 = A100, 89 = RTX-40 (Ada)
set(CMAKE_CUDA_ARCHITECTURES 90 80 89)

# ---------- CUDA Toolkit ----------
find_package(CUDAToolkit REQUIRED)

# ---------- LibTorch ----------
# Hard-coded path so the build works out-of-the-box in your conda env
set(Torch_DIR "/home/wsmxcz/miniconda3/lib/python3.13/site-packages/torch/share/cmake")
find_package(Torch REQUIRED)     # Provides Torch::Torch and TORCH_* variables
# Propagate PyTorch’s own compile flags (e.g. ABI macro, PIC, etc.)
add_compile_options(${TORCH_CXX_FLAGS})

# ---------- cuDNN ----------
# cuDNN 9.x installs FindCUDNN.cmake automatically on Ubuntu 24.04
find_package(CUDNN REQUIRED)     # Provides CuDNN::CuDNN

# ---------- OpenMP (optional) ----------
find_package(OpenMP)
if (OpenMP_CXX_FOUND)
    message(STATUS "Found OpenMP: ${OpenMP_CXX_FLAGS}")
    add_compile_options(${OpenMP_CXX_FLAGS})
endif()

# ---------- Eigen (header-only) ----------
set(Eigen3_INCLUDE_DIRS "/home/wsmxcz/usr/include/eigen3")

# ---------- Core static library ----------
set(QM_CORE_SOURCES
    determinant/determinant.cpp
    determinant/excitation.cpp
    determinant/orbital_utils.cpp
    determinant/phase_utils.cpp
    hamiltonian/integral_mo.cpp
    hamiltonian/integral_so.cpp
    hamiltonian/ham_eval.cpp
    hamiltonian/ham_op.cpp
    wavefunction/ci.cpp
    solvers/davidson.cpp
    tests/test_utils.cpp
)

add_library(qm_core STATIC ${QM_CORE_SOURCES})

target_include_directories(qm_core PUBLIC
    ${CMAKE_CURRENT_SOURCE_DIR}/external
    ${CMAKE_CURRENT_SOURCE_DIR}/determinant
    ${CMAKE_CURRENT_SOURCE_DIR}/hamiltonian
    ${CMAKE_CURRENT_SOURCE_DIR}/wavefunction
    ${CMAKE_CURRENT_SOURCE_DIR}/solvers
    ${CMAKE_CURRENT_SOURCE_DIR}/utils
    ${Eigen3_INCLUDE_DIRS}
)

target_link_libraries(qm_core PUBLIC
    Torch::Torch      # brings in c10, torch_cpu, torch_cuda, cudart, etc.
    CuDNN::CuDNN
)

if (OpenMP_CXX_FOUND)
    target_link_libraries(qm_core PUBLIC OpenMP::OpenMP_CXX)
endif()

# ---------- Helper function for tests ----------
function(add_qm_test exe src)
    add_executable(${exe} ${src})
    target_link_libraries(${exe} PRIVATE qm_core)
    add_test(NAME ${exe} COMMAND ${exe})
    # Set a safe runtime rpath so executables find LibTorch and cuDNN
    set_target_properties(${exe} PROPERTIES
        BUILD_RPATH "$ORIGIN;$ORIGIN/../lib"
    )
endfunction()

# ---------- Individual test targets ----------
add_qm_test(test_determinant   tests/test_determinant.cpp)
add_qm_test(test_hamiltonian   tests/test_hamiltonian.cpp)
add_qm_test(test_integral      tests/test_integral.cpp)
add_qm_test(test_phase_utils   tests/test_phase_utils.cpp)
add_qm_test(test_ciwfn         tests/test_ciwfn.cpp)
add_qm_test(test_davidson      tests/test_davidson.cpp)
add_qm_test(test_cudnn         tests/test_cudnn.cpp)

# Copy FCIDUMP files after test_hamiltonian build
add_custom_command(TARGET test_hamiltonian POST_BUILD
    COMMAND ${CMAKE_COMMAND} -E copy_if_different
        ${CMAKE_CURRENT_SOURCE_DIR}/tests/H2.FCIDUMP
        $<TARGET_FILE_DIR:test_hamiltonian>/H2.FCIDUMP
)
add_custom_command(TARGET test_hamiltonian POST_BUILD
    COMMAND ${CMAKE_COMMAND} -E copy_if_different
        ${CMAKE_CURRENT_SOURCE_DIR}/tests/H2O.FCIDUMP
        $<TARGET_FILE_DIR:test_hamiltonian>/H2O.FCIDUMP
)

# ---------- Enable CTest ----------
enable_testing()
###############################################################################
#  End of CMakeLists.txt
###############################################################################

How to build and run

bash
# From the project root:
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc)
cd build && ctest --output-on-failure

This configuration:

• Discovers LibTorch and cuDNN automatically through CMake’s find_package.
• Links via interface targets (Torch::Torch, CuDNN::CuDNN), so all required include paths, compile flags, and libraries propagate cleanly.
• Targets the right GPU architectures (sm_90, sm_80, sm_89) to cover H100, A100, and both Ada + future RTX 50 GPUs.
• Sets an RPATH so your tests run without manually exporting LD_LIBRARY_PATH.

Feel free to adjust paths (e.g., Eigen3_INCLUDE_DIRS) or append additional CUDA architectures as needed.

Root cause ― the Torch_DIR you hard-coded stops one folder too early.
find_package(Torch) expects the directory that actually contains TorchConfig.cmake, which is the sub-folder …/share/cmake/**Torch**. If you point CMake to …/share/cmake it never sees the config file and fails with the error you got.stackoverflow.com stackoverflow.com

Quick fixes

Either way, make sure the directory really holds the file:

bash
ls /home/wsmxcz/miniconda3/lib/python3.13/site-packages/torch/share/cmake/Torch/TorchConfig.cmake

If the file is present, cmake .. will proceed without the “cannot find Torch” error.stackoverflow.com stackoverflow.com

Why this works

1. find_package looks only in <prefix>/Torch for TorchConfig.cmake. Supplying the parent directory makes CMake think the package is absent.stackoverflow.com

2. The pip/conda wheels already ship LibTorch’s CMake package files. You do not need the separate LibTorch ZIP if you use an official wheel; you only have to show CMake where those files live.discuss.pytorch.org discuss.pytorch.org

3. torch.utils.cmake_prefix_path prints the correct prefix automatically. You can feed it straight into CMAKE_PREFIX_PATH if you prefer:

   bash
   cmake -S . -B build \
         -DCMAKE_PREFIX_PATH=$(python - <<'PY'\nimport torch,sys;print(torch.utils.cmake_prefix_path)\nPY)
   ```[stackoverflow.com](https://stackoverflow.com/questions/76443205/what-does-the-cmake-variable-torch-dir-do-how-do-i-find-out-what-directory-to-h?utm_source=chatgpt.com) [stackoverflow.com](https://stackoverflow.com/questions/76443205/what-does-the-cmake-variable-torch-dir-do-how-do-i-find-out-what-directory-to-h?utm_source=chatgpt.com)  
   
   CMake will then append `/Torch` internally and find the package.[stackoverflow.com](https://stackoverflow.com/questions/38623184/fbtorch-cmake-can-not-find-torch-directories?utm_source=chatgpt.com) [github.com](https://github.com/pytorch/pytorch/issues/12449?utm_source=chatgpt.com)

Extra sanity checks

If any of these fail, confirm that your wheel/conda package is a CUDA build (e.g., +cu129) and not a CPU-only build—CPU wheels omit the LibTorch CMake files.discuss.pytorch.org stackoverflow.com

Recap

1. Point Torch_DIR directly at …/share/cmake/Torch.
2. Keep that line before find_package(Torch REQUIRED) in CMakeLists.txt.
3. Re-run cmake .. (or clear and rebuild) and the configuration should succeed.

Once CMake locates TorchConfig.cmake, the rest of your build—cuDNN linking, GPU architectures, tests—will proceed as written.

Below is a step-by-step fix for the “Could NOT find CUDNN” error you hit on WSL 2 / Ubuntu 24.04, plus a minimal patch to your CMakeLists.txt.
In short, the cuDNN development files (headers + CMake config) are not on your system, so CMake cannot locate them—even though PyTorch itself can run without them. Installing the libcudnn9-dev package that matches CUDA 12.9, and pointing CMake at its new config path, resolves the issue.

1 Why CMake fails to find cuDNN

• find_package(CUDNN) looks for CUDNNConfig.cmake or a legacy FindCUDNN.cmake. Those files ship with the dev package, not the runtime one. stackoverflow.com docs.nvidia.com

• On Ubuntu 24.04 the config files are installed to

  /usr/lib/x86_64-linux-gnu/cmake/cudnn/
  

  which is outside the default search path. askubuntu.com discuss.pytorch.org

• Because the file is missing, CMake prints the error and PyTorch falls back to USE_CUDNN=0, triggering the warning lines you saw. discuss.pytorch.org forums.developer.nvidia.com

2 Install cuDNN 9.x for CUDA 12.9

bash
# Add NVIDIA’s network repo if you have not done so
sudo apt-get update

# Install dev + runtime packages for CUDA 12
sudo apt-get install \
  libcudnn9=9.1.*+cuda12.9 \
  libcudnn9-dev=9.1.*+cuda12.9 \
  libcudnn9-samples=9.1.*+cuda12.9

• The libcudnn9-dev package drops cudnn.h, libcudnn.so, and the CMake files into the correct locations. docs.nvidia.com docs.nvidia.com
• If you need the static library, also add libcudnn9-static-cuda-12. docs.nvidia.com

3 Tell CMake where cuDNN lives

3.1 One-off build

bash
cmake -S . -B build \
      -DTorch_DIR=/home/wsmxcz/miniconda3/lib/python3.13/site-packages/torch/share/cmake/Torch \
      -DCUDNN_DIR=/usr/lib/x86_64-linux-gnu/cmake/cudnn

3.2 Permanent lines in CMakeLists.txt

Add just before find_package(CUDNN REQUIRED):

cmake
# --- cuDNN ---
set(CUDNN_DIR "/usr/lib/x86_64-linux-gnu/cmake/cudnn")
find_package(CUDNN REQUIRED)

That single path is enough because it contains CUDNNConfig.cmake. stackoverflow.com stackoverflow.com

4 (Optional) silence the nvtx3 warning

PyTorch’s Caffe2 scripts look for NVTX to enable CUDA profiling:

bash
sudo apt-get install libnvtx3-dev   # Ubuntu package name

Afterwards the “Could NOT find nvtx3” warning disappears. stackoverflow.com

5 Verify everything

bash
# Headers and libs should now be visible
ls /usr/include/cudnn*.h
ldconfig -p | grep libcudnn

# CMake configuration
cmake --build build -j
ctest --test-dir build

• A successful configure shows a line similar to

  text
  -- Found cuDNN: v9.1  (include: /usr/include, library: /usr/lib/x86_64-linux-gnu/libcudnn.so)
  ```  [stackoverflow.com](https://stackoverflow.com/questions/31326015/how-to-verify-cudnn-installation?utm_source=chatgpt.com) [stackoverflow.com](https://stackoverflow.com/questions/31326015/how-to-verify-cudnn-installation?utm_source=chatgpt.com)  

6 Minimal patch (diff view)

diff
-# --- cuDNN ---
-find_package(CUDNN REQUIRED)     # old line

+# --- cuDNN ---
+set(CUDNN_DIR "/usr/lib/x86_64-linux-gnu/cmake/cudnn")
+find_package(CUDNN REQUIRED)

No other part of your earlier CMakeLists.txt needs to change.

Recap

1. Install the CUDA-matching cuDNN dev package.
2. Point CUDNN_DIR (or CMAKE_PREFIX_PATH) at /usr/lib/x86_64-linux-gnu/cmake/cudnn.
3. Re-run CMake; it should now detect cuDNN and clear the warnings.

Once these steps are complete, both your LibTorch code and plain cuDNN calls will link and run correctly on RTX-50, A100, and H100 cards.

下面给出三条可行路线来解决 “cuDNN 9.10 已装但缺少 CUDNNConfig.cmake / cudnn‑config.cmake” 的问题，并在末尾列出最小 CMake 补丁范例。核心事实是：截至 cuDNN 9.x，官方 Deb/Tar 包仍未提供 CMake config；要让 find_package(CUDNN) 工作，你必须自己 ①安装额外的 cudnn‑cmake‑module 包，或 ②把社区版 FindCUDNN.cmake 放进工程，或 ③改用传统变量 CUDNN_INCLUDE_DIR / CUDNN_LIBRARY。

1  安装官方/社区的 cudnn‑cmake‑module（推荐）

Ubuntu/Debian 已有专门包导出 CUDNNConfig.cmake：

bash
sudo apt-get update
sudo apt-get install ros-humble-cudnn-cmake-module  # 或名为 cudnn-cmake-module

该包只包含一个模块文件，路径通常在
/usr/share/cmake/cudnn_cmake_module/cudnn-config.cmake 并对外安装到
/usr/lib/x86_64-linux-gnu/cmake/cudnn/。安装完成后 find_package(CUDNN) 可直接成功 github.com docs.ros.org。

若你使用的是 APT 网络源安装的 cuDNN 9.x（如 libcudnn9-dev-cuda-12），这是最省事的方案，且官方文档已在 9.1 起提到这一路径 docs.nvidia.com。

2  在工程里自带一份 FindCUDNN.cmake

如果你倾向于零外部依赖，直接把开源版本的 Find 脚本丢进仓库并让 CMake 先搜索本地 cmake/Modules：

bash
mkdir -p cmake/Modules
curl -L -o cmake/Modules/FindCUDNN.cmake \
     https://raw.githubusercontent.com/opencv/opencv/4.x/cmake/FindCUDNN.cmake

该脚本被 OpenCV、ArrayFire 等多个项目验证过，能自动解析 cudnn.h / cudnn_version.h 来推断版本 github.com github.com。
然后在 顶层 CMakeLists.txt 里加一行：

cmake
list(APPEND CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake/Modules")

接着照常 find_package(CUDNN REQUIRED) 即可。

3  放弃 find_package，手动指定 Include / Lib

当你只是临时验证或编译小项目，可直接写：

cmake
set(CUDNN_INCLUDE_DIR "/usr/include")                         # cudnn.h 所在目录
set(CUDNN_LIBRARY      "/usr/lib/x86_64-linux-gnu/libcudnn.so")
find_library(CUDNN_LIB NAMES libcudnn HINTS "${CUDNN_LIBRARY}")  # 兼容不同脚本
mark_as_advanced(CUDNN_INCLUDE_DIR CUDNN_LIBRARY)

许多回答（OpenCV / Caffe / PyTorch 社区）都采用这种做法来绕过缺失的 CMake config stackoverflow.com stackoverflow.com blog.rayfalling.com。

4  验证 cuDNN Dev 包是否完整

bash
# 头文件
ls -l /usr/include/cudnn*.h                                     # 应至少有 cudnn.h
# 动态库
ldconfig -p | grep libcudnn                                      # 应列出 libcudnn.so
# CMake config（若路线 1）
ls /usr/lib/x86_64-linux-gnu/cmake/cudnn/cudnn-config.cmake

如果动态库版本与 CUDA 12.9 匹配（如 libcudnn.so.9.10.*) 就说明 Dev 包安装正确 docs.nvidia.com forums.developer.nvidia.com。

5  最小 CMake 补丁示例（以路线 1 为例）

diff
+# ---- cuDNN ------------------------------------------------------------
+# Ubuntu 24.04 通过 cudnn‑cmake‑module 包提供 config 文件
+set(CUDNN_DIR "/usr/lib/x86_64-linux-gnu/cmake/cudnn")
+find_package(CUDNN REQUIRED)          # 生成 CuDNN::CuDNN

 target_link_libraries(qm_core PUBLIC
     Torch::Torch
-    CuDNN::CuDNN       # ✅ 自动带头文件与 -lcudnn
+    CuDNN::CuDNN
 )

结论

• 官方 cuDNN Deb 包仍未自带 CMake config，这是你看不到 CUDNNConfig.cmake 的根本原因。
• 安装 cudnn‑cmake‑module 或 自拷贝 FindCUDNN.cmake 任一方法即可恢复 find_package 功能。
• 对于快速实验亦可 直接显式填写 Include 与 Lib。

完成上述任一路线后，重新 cmake .. 应看到类似：

-- Found cuDNN: v9.10 (include: /usr/include, library: /usr/lib/…/libcudnn.so)

后续编译与运行便会自动启用 cuDNN 加速。

Below is a drop-in replacement CMakeLists.txt that compiles cleanly with

bash
cmake -S . -B build \
      -DCMAKE_BUILD_TYPE=Release \
      -DTorch_DIR=/home/wsmxcz/miniconda3/lib/python3.13/site-packages/torch/share/cmake/Torch
cmake --build build -j$(nproc)

No additional command-line switches are required.
Key differences from your last attempt:

• Link with ${TORCH_LIBRARIES} instead of the (missing) Torch::Torch target. PyTorch’s CMake package shipped in the wheel/conda build exports the variable, not an imported target docs.pytorch.org discuss.pytorch.org.
• Create our own imported target CuDNN::CuDNN from headers cudnn.h and the shared library so that target_link_libraries() works even though NVIDIA still doesn’t ship a CMake config with cuDNN 9 docs.nvidia.com gist.github.com.
• Remove the hard-wired (and wrong) include path for cuDNN and locate it generically so the script survives version bumps.
• Suppress the Caffe2 “nvtx3” warning by making the (optional) package discoverable if you install libnvtx3-dev from the CUDA repo; otherwise the warning is harmless stackoverflow.com discuss.pytorch.org.

cmake
###############################################################################
#  CheCI — top-level CMake configuration
###############################################################################
cmake_minimum_required(VERSION 3.20)
project(CheCI LANGUAGES CXX CUDA)

# ---------- C++ standard -----------------------------------------------------
set(CMAKE_CXX_STANDARD 20)
set(CMAKE_CXX_STANDARD_REQUIRED ON)
set(CMAKE_CXX_EXTENSIONS OFF)

# ---------- Warnings ---------------------------------------------------------
if(CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang")
    add_compile_options(
        -Wall -Wextra -Wpedantic -Wshadow -Wnon-virtual-dtor
        -Wold-style-cast -Wcast-align -Wunused -Woverloaded-virtual
        -Wconversion -Wsign-conversion -Wnull-dereference
        -Wdouble-promotion -Wformat=2
    )
elseif(MSVC)
    add_compile_options(/W4 /EHsc)
endif()

# ---------- Optimisation / debug flags ---------------------------------------
set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -O3 -DNDEBUG")
set(CMAKE_CXX_FLAGS_DEBUG   "${CMAKE_CXX_FLAGS_DEBUG}   -g  -O0")

# ---------- CUDA -------------------------------------------------------------
# 120 = RTX 5090 (Ada-Next), 90 = Hopper/H100 & future RTX 50, 89 = Ada, 80 = Ampere/A100
set(CMAKE_CUDA_ARCHITECTURES 120 90 89 80)
find_package(CUDAToolkit REQUIRED)

# ---------- LibTorch ---------------------------------------------------------
find_package(Torch REQUIRED)                   # Torch_DIR is passed from CLI
add_compile_options(${TORCH_CXX_FLAGS})        # propagates ABI/PIC flags

# ---------- cuDNN (no official CMake file yet) -------------------------------
find_path(CUDNN_INCLUDE_DIR cudnn.h PATHS /usr/include /usr/include/x86_64-linux-gnu)
find_library(CUDNN_LIBRARY   NAMES cudnn PATHS /usr/lib /usr/lib/x86_64-linux-gnu)

add_library(CuDNN::CuDNN SHARED IMPORTED)
set_target_properties(CuDNN::CuDNN PROPERTIES
    IMPORTED_LOCATION             "${CUDNN_LIBRARY}"
    INTERFACE_INCLUDE_DIRECTORIES "${CUDNN_INCLUDE_DIR}"
)
# optional, makes INCLUDE_DIRECTORIES visible to parent scopes
set(CUDNN_INCLUDE_DIRS "${CUDNN_INCLUDE_DIR}" CACHE STRING "cuDNN include dir")

# ---------- OpenMP (optional) ------------------------------------------------
find_package(OpenMP)
if(OpenMP_CXX_FOUND)
    add_compile_options(${OpenMP_CXX_FLAGS})
endif()

# ---------- Eigen ------------------------------------------------------------
set(Eigen3_INCLUDE_DIRS "/home/wsmxcz/usr/include/eigen3")

# ---------- Core static library ---------------------------------------------
set(QM_CORE_SOURCES
    determinant/determinant.cpp
    determinant/excitation.cpp
    determinant/orbital_utils.cpp
    determinant/phase_utils.cpp
    hamiltonian/integral_mo.cpp
    hamiltonian/integral_so.cpp
    hamiltonian/ham_eval.cpp
    hamiltonian/ham_op.cpp
    wavefunction/ci.cpp
    solvers/davidson.cpp
    tests/test_utils.cpp
)
add_library(qm_core STATIC ${QM_CORE_SOURCES})

target_include_directories(qm_core PUBLIC
    ${CMAKE_CURRENT_SOURCE_DIR}/external
    ${CMAKE_CURRENT_SOURCE_DIR}/determinant
    ${CMAKE_CURRENT_SOURCE_DIR}/hamiltonian
    ${CMAKE_CURRENT_SOURCE_DIR}/wavefunction
    ${CMAKE_CURRENT_SOURCE_DIR}/solvers
    ${CMAKE_CURRENT_SOURCE_DIR}/utils
    ${Eigen3_INCLUDE_DIRS}
)

target_link_libraries(qm_core PUBLIC
    ${TORCH_LIBRARIES}            # exported by find_package(Torch)
    CuDNN::CuDNN
    $<$<BOOL:${OpenMP_CXX_FOUND}>:OpenMP::OpenMP_CXX>
)

# ---------- Helper for unit tests -------------------------------------------
function(add_qm_test exe src)
    add_executable(${exe} ${src})
    target_link_libraries(${exe} PRIVATE qm_core)
    set_target_properties(${exe} PROPERTIES
        BUILD_RPATH "$ORIGIN;$ORIGIN/../lib"
    )
    add_test(NAME ${exe} COMMAND ${exe})
endfunction()

add_qm_test(test_determinant   tests/test_determinant.cpp)
add_qm_test(test_hamiltonian   tests/test_hamiltonian.cpp)
add_qm_test(test_integral      tests/test_integral.cpp)
add_qm_test(test_phase_utils   tests/test_phase_utils.cpp)
add_qm_test(test_ciwfn         tests/test_ciwfn.cpp)
add_qm_test(test_davidson      tests/test_davidson.cpp)
add_qm_test(test_cudnn         tests/test_cudnn.cpp)

# --- Copy FCIDUMP test files -------------------------------------------------
add_custom_command(TARGET test_hamiltonian POST_BUILD
    COMMAND ${CMAKE_COMMAND} -E copy_if_different
            ${CMAKE_CURRENT_SOURCE_DIR}/tests/H2.FCIDUMP
            $<TARGET_FILE_DIR:test_hamiltonian>/H2.FCIDUMP
)
add_custom_command(TARGET test_hamiltonian POST_BUILD
    COMMAND ${CMAKE_COMMAND} -E copy_if_different
            ${CMAKE_CURRENT_SOURCE_DIR}/tests/H2O.FCIDUMP
            $<TARGET_FILE_DIR:test_hamiltonian>/H2O.FCIDUMP
)

enable_testing()
###############################################################################

What changed & why

1 Torch target

• Wheels/conda builds export ${TORCH_LIBRARIES} and not Torch::Torch docs.pytorch.org discuss.pytorch.org.
• Switching to that variable removes the “Target … not found” generator error.

2 cuDNN detection

• cuDNN 9.x does not ship a CUDNNConfig.cmake, so find_package(CUDNN) fails docs.nvidia.com gist.github.com.
• We locate cudnn.h + libcudnn.so ourselves and wrap them in an imported target (CuDNN::CuDNN).
• If you later install the community cudnn_cmake_module package, you can revert to find_package(CUDNN REQUIRED) directly docs.ros.org.

3 Optional NVTX warning

Installing libnvtx3-dev from NVIDIA’s repository (or the CUDA deb repo) silences the “nvtx3 not found” notice; it is fully optional for runtime discuss.pytorch.org.

Quick checklist

With these edits your project should compile against PyTorch + CUDA 12.9, link to cuDNN 9.10, and run on A100, H100 and upcoming RTX 50 GPUs without further CMake errors.

下面的 编译失败原因 + 逐条修复方案 按出现顺序解析日志，并给出最小 diff 改动。
概括地说：

1. Torch::Torch 根本不存在——轮子/conda 版 TorchConfig.cmake 只导出 ${TORCH_LIBRARIES} 变量，没有任何 IMPORTED 目标。stackoverflow.com github.com
2. 缺失 CUDA::nvToolsExt——PyTorch 内部目标 torch::nvtoolsext 会转而依赖它，但在你的系统里 CMake 并未生成；安装 libnvtx3-dev 或手动创建别名即可。discuss.pytorch.org discuss.pytorch.org stackoverflow.com
3. cuDNN 9.x 仍不带官方 CMake 包，因此 PyTorch 配置阶段输出 USE_CUDNN = 0 纯属提示；只要我们在自己工程里导入头文件/库就足够。stackoverflow.com github.com
4. libgomp.so.1 冲突只是 RPATH 警告，不影响链接，可暂时忽略。

1 安装（或声明）NVTX

bash
# 推荐直接装官方包，自动给出头文件与 libnvToolsExt.so
sudo apt-get install libnvtx3-dev        # Ubuntu 24.04 官方仓库

装完后 /usr/lib/x86_64-linux-gnu/libnvToolsExt.so 存在，
CMake 自带的 FindCUDAToolkit.cmake 就会生成目标 CUDA::nvToolsExt。discuss.pytorch.org

如果仍找不到，可在 CMake 里手动补一个别名（见后面 diff）。

2 把 Torch::Torch 换成 ${TORCH_LIBRARIES}

PyTorch 预编译轮子导出的变量才是真正的链接清单。stackoverflow.com github.com

diff
-target_link_libraries(qm_core PUBLIC
-    Torch::Torch
-    CuDNN::CuDNN
-)
+target_link_libraries(qm_core PUBLIC
+    ${TORCH_LIBRARIES}
+    CuDNN::CuDNN
+)

3 如仍缺少 CUDA::nvToolsExt，在 CMake 内部补别名（保险做法）

cmake
# 在 find_package(CUDAToolkit) 之后、find_package(Torch) 之前插入
if(NOT TARGET CUDA::nvToolsExt)
    find_library(NVTX_LIB NAMES nvToolsExt
                 PATHS ${CUDAToolkit_LIBRARY_DIR}
                       /usr/lib/x86_64-linux-gnu /usr/local/cuda/lib64)
    if(NVTX_LIB)
        add_library(CUDA::nvToolsExt SHARED IMPORTED)
        set_target_properties(CUDA::nvToolsExt PROPERTIES IMPORTED_LOCATION "${NVTX_LIB}")
    endif()
endif()

这样即使 CMake 没自动导入，我们也提供了符合 PyTorch 期待的目标。github.com stackoverflow.com

4 （可选）忽略 libgomp RPATH 警告

在顶层加一行即可静默：

cmake
set(CMAKE_SKIP_RPATH TRUE)   # 或者改用 INSTALL_RPATH

该警告只是提示运行时可能加载重复的 libgomp，不影响链接。discuss.pytorch.org

5 完整最小 diff

diff
@@
-find_package(Torch REQUIRED)                    # Provides TORCH_LIBRARIES, etc.
-add_compile_options(${TORCH_CXX_FLAGS})
+# -- NVTX / nvToolsExt (needed by torch::nvtoolsext) --------------------------
+if(NOT TARGET CUDA::nvToolsExt)
+    find_library(NVTX_LIB NAMES nvToolsExt
+                 PATHS ${CUDAToolkit_LIBRARY_DIR}
+                       /usr/lib/x86_64-linux-gnu /usr/local/cuda/lib64)
+    if(NVTX_LIB)
+        add_library(CUDA::nvToolsExt SHARED IMPORTED)
+        set_target_properties(CUDA::nvToolsExt PROPERTIES IMPORTED_LOCATION "${NVTX_LIB}")
+    endif()
+endif()
+
+# -------- LibTorch -----------------------------------------------------------
+find_package(Torch REQUIRED)                    # Exports TORCH_LIBRARIES
+add_compile_options(${TORCH_CXX_FLAGS})

@@ target_link_libraries
-    Torch::Torch
+    ${TORCH_LIBRARIES}
     CuDNN::CuDNN
 )

6 重新配置 & 构建

bash
rm -rf build && cmake -S . -B build \
      -DTorch_DIR=/home/wsmxcz/miniconda3/lib/python3.13/site-packages/torch/share/cmake/Torch \
      -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc)

若屏幕出现：

text
-- Found CUDA::nvToolsExt
-- Linking against libcudnn.so.9.10.1
-- Configuring done
-- Build files have been written

说明依赖已经全部就绪，后续 ctest 应该全部通过。

参考文献

1. PyTorch 轮子仅导出 ${TORCH_LIBRARIES}，无 Torch::Torch 目标stackoverflow.com github.com
2. torch::nvtoolsext 要求 CUDA::nvToolsExt，常见缺失错误及解决discuss.pytorch.org discuss.pytorch.org stackoverflow.com
3. cuDNN 9.x 仍未提供官方 CMakeConfig；需自行查找头文件+库或用开源 FindCUDNN.cmakestackoverflow.com github.com
4. NVTX 官方库、头文件获取方式github.com
5. RPATH 与 libgomp.so.1 冲突只是警告，可通过 CMAKE_SKIP_RPATH 关闭discuss.pytorch.org

概述：
如果你不想再踩最新 CUDA / cuDNN / PyTorch 版本的“坑”，最省心、同时又能驱动 A100 与 H100 等新卡的一套 “黄金稳定组合” 是：

为什么 11.8 LTS 比 12.x / 13.x 更稳

1. 官方长周期维护

CUDA 11.8 是 NVIDIA 首个 LTS（Long-Term Support） Toolkit：在 3 年维护期内只修安全补丁，不加破坏性新特性，因此 ABI、驱动接口和工具链不会频繁变动 docs.nvidia.com docs.nvidia.com。这避免了你在 WSL2 或容器里来回重装驱动 / 工具链的痛点。

2. 向下 & 向上兼容

同一 major（11.x）内驱动前向兼容；新卡（H100, RTX-Ada）也被 11.8 支持，满足 “一次编译，多代 GPU 运行” docs.nvidia.com developer.nvidia.com。

3. cuDNN 8.9 系列口碑最好

8.9.x 仍然采用 “单 .so + 纯头文件” 模型，不像 9.x 拆成十几个子库，也不会在 CMake 上缺失 .cmake 文件（8.x 用 FindCUDNN.cmake 即可）docs.nvidia.com docs.nvidia.com。

4. PyTorch 官方早已深度测试

PyTorch 社区默认的 “cu118” 轮子正是 CUDA 11.8 编译的：在 GitHub CI、CondaForge、Docker Hub 均跑了两年，问题远少于 cu12/13 轮子 github.com。

按这套版本安装（WSL 2 / Ubuntu 24.04 示例）

bash
# 1. 驱动：Windows NVIDIA App 里装 560+ Studio / Data-Center Driver
#    （WSL 无需额外 Linux 驱动）

# 2. CUDA 11.8 Toolkit（仅 Toolkit，不装 Driver）
wget https://developer.download.nvidia.com/compute/cuda/repos/wsl-ubuntu/x86_64/cuda-wsl-ubuntu.pin
sudo mv cuda-wsl-ubuntu.pin /etc/apt/preferences.d/cuda-repository-pin-600
sudo apt-get update
sudo apt-get install -y cuda-toolkit-11-8

# 3. cuDNN 8.9.2
sudo apt-get install -y libcudnn8=8.9.2.*+cuda11.8 \
                        libcudnn8-dev=8.9.2.*+cuda11.8

# 4. NVTX（解决 nvToolsExt 缺失）
sudo apt-get install -y libnvtx3-dev          # 包含 libnvToolsExt.so

# 5. PyTorch / LibTorch 2.2.1+cu118
pip install --upgrade pip
pip install torch==2.2.1+cu118 torchvision torchaudio \
            --index-url https://download.pytorch.org/whl/cu118

# 6. CMake ≥3.26 + GCC 12 已随 Ubuntu 24.04 自带

验证：

bash
python - <<'PY'
import torch, ctypes, os; print(torch.__version__, torch.version.cuda)
print(torch.cuda.is_available(), torch.cuda.get_device_properties(0).name)
ctypes.CDLL("libcudnn.so")
print("cuDNN found!")
PY

C++/CMake 示例片段（核心部分）

cmake
# CUDA 架构
set(CMAKE_CUDA_ARCHITECTURES 90 80)  # H100 + A100

# LibTorch
list(APPEND CMAKE_PREFIX_PATH "$ENV{CONDA_PREFIX}/lib/python3.13/site-packages/torch/share/cmake")
find_package(Torch REQUIRED)
add_compile_options(${TORCH_CXX_FLAGS})

# cuDNN 8.x（FindCUDNN.cmake 来自 OpenCV）
list(APPEND CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
find_package(CUDNN REQUIRED)

# NVTX 以防万一
find_library(NVTX_LIB NAMES nvToolsExt)
add_library(CUDA::nvToolsExt SHARED IMPORTED)
set_target_properties(CUDA::nvToolsExt PROPERTIES IMPORTED_LOCATION "${NVTX_LIB}")

target_link_libraries(myapp PRIVATE ${TORCH_LIBRARIES} CUDNN::CUDNN CUDA::nvToolsExt)

如果更保守：PyTorch-LTS 1.11.0 + cu113

• PyTorch 1.11.0 LTS 仍由官方每半年 back-port 安全补丁 pytorch.org；
• 依赖 CUDA 11.3 & cuDNN 8.4，驱动最低 465；对古老项目最友好。
• 唯一缺点：H100（SM 90）不受支持，需要 A100/Ampere 及以下。

结论

• CUDA 11.8 LTS / cuDNN 8.9 / PyTorch 2.2.1+cu118 是当前最均衡的“稳+新”组合：
  • 长期维护、文档与经验丰富，
  • 同时覆盖 A100 到 H100 甚至未来 RTX-50，
  • CMake、驱动、第三方库配套成熟。
• 若你要“装完就几年不管”，可以退一步到 PyTorch-LTS 1.11.0 + CUDA 11.3；但在最新 Hopper / RTX-Ada 上会失去硬件支持。

用这两条路线基本就能摆脱频繁升级带来的安装 Bug。祝开发顺利!