memo: LibTorch | Debug

Table of contents

C++ Extensions - Docs

Install LibTorch

libtorch is the C++ API for PyTorch. Guide: INSTALLING C++ DISTRIBUTIONS OF PYTORCH

  1. Download the binary files with modified URL for specified version:

    1
2
3
4
5
6
7

    # Download here (cxx11 ABI):
wget https://download.pytorch.org/libtorch/cu116/libtorch-cxx11-abi-shared-with-deps-1.12.1%2Bcu116.zip

unzip libtorch-cxx11-abi-shared-with-deps-1.12.1+cu116.zip

# Optional
sudo mv libtorch/ /usr/local/

  2. Create folder “example-app”, and write code into “example-app.cpp”

    1
2
3
4
5
6
7
8

    #include <torch/torch.h>
#include <iostream>

int main(){
    torch::Tensor tensor = torch::rand({2,3});
    std::cout << tensor << std::endl;
    std::cout << "size of Tensor type: "<< sizeof(torch::Tensor) << std::endl;
}

  3. Set dependence for LibTorch via CMakeLists.txt:

    1
2
3
4
5
6
7
8
9

    cmake_minimum_required(VERSION 3.18 FATAL_ERROR)
project(MyLibTorchApp)    # name

find_package(Torch REQUIRED)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${TORCH_CXX_FLAGS}")

add_executable(MyLibTorchApp main.cpp)
target_link_libraries(MyLibTorchApp "${TORCH_LIBRARIES}")
set_property(TARGET MyLibTorchApp PROPERTY CXX_STANDARD 17)

  4. Build (externally) inside the dir “example-app/build”

    1
2
3
4

    # mkdir build
# cd build
cmake -DCMAKE_PREFIX_PATH=/absolute/path/to/libtorch ..
cmake --build . --config Release

  5. (Optional) For C/C++ extension, set up includePath in “c_cpp_properties.json” for correct intellisense:

     1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20

    {
    "configurations": [
        {
            "name": "Linux",
            "includePath": [
                "${workspaceFolder}/**",
                "/usr/local/libtorch",
                "/home/yi/anaconda3/envs/AIkui/include/python3.10",
                "/home/yi/anaconda3/envs/AIkui/lib/python3.10/site-packages/torch/include",
                "/home/yi/anaconda3/envs/AIkui/lib/python3.10/site-packages/torch/include/torch/csrc/api/include"
            ],
            "defines": [],
            "compilerPath": "/usr/bin/g++",
            "cStandard": "c17",
            "cppStandard": "gnu++14",
            "intelliSenseMode": "linux-gcc-x64"
        }
    ],
    "version": 4
}

Debug LibTorch

(2023-11-12)

launch.json

Modify launch.json in .vscode to debug C/C++ binary file:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14

{
    "version": "0.2.0",
    "configurations": [ 
        {
            "name": "LibTorch Launch",
            "type": "cppdbg",
            "request": "launch",
            "program": "${fileDirname}/build/MyLibTorchApp",
            "args": [],
            "preLaunchTask": "Build with cmake", 
            "cwd": "${workspaceFolder}",
        }
    ]
}

task.json

Set up “tasks.json” for building project with cmake:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44

{
	"version": "2.0.0",
	"tasks": [
	  {
		"label": "cmake-configure",
		"type": "shell",
		"command": "cmake",
		"args": [
		  "-DCMAKE_BUILD_TYPE=Debug",
		  "-DCMAKE_PREFIX_PATH=/absolute/path/to/libtorch",
		  "..",
		],
		"options": {
		  "cwd": "${workspaceFolder}/build"  // Set the build directory
		},
		"group": {
		  "kind": "build",
		  "isDefault": true
		}
	  },
	  {
		"label": "cmake-build",
		"type": "shell",
		"command": "cmake",
		"args": [
		  "--build",
		  ".",
		  "--config",
		  "Debug"
		],
		"options": {
		  "cwd": "${workspaceFolder}/build"  // Set the build directory
		},
		"group": {
		  "kind": "build",
		  "isDefault": true
		}
	  },
	  {
		"label": "Build with cmake",
		"dependsOn": ["cmake-configure", "cmake-build"]
	  }
	]
}

  1. Multiple tasks can be combined to one, which then is able tobe passed to preLaunchTask in “launch.json”. (Can’t pass 2 tasks at once.)

    Integrate with External Tools via Tasks - Docs

  2. The above 2 tasks both are custom tasks executed in shell.

  3. CMake requires two steps, so 2 tasks are needed.

  4. The tag -DCMAKE_BUILD_TYPE=Debug is needed for generating debug info. Otherwise, the breakpoints won’t take effect:

    1

    Module containing this breakpoint has not yet loaded or the breakpoint address could not be obtained.

  5. However, sometimes, camke build failed:

    Expand the error 
     1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54

     *  Executing task: cmake --build . --config Debug 

-- Caffe2: CUDA detected: 11.6
-- Caffe2: CUDA nvcc is: /usr/local/cuda-11.6/bin/nvcc
-- Caffe2: CUDA toolkit directory: /usr/local/cuda-11.6
CMake Error at /usr/local/libtorch/share/cmake/Caffe2/public/cuda.cmake:88 (message):
  Caffe2: Couldn't determine version from header: Change Dir:
  '/home/yi/Downloads/example-app/build/CMakeFiles/CMakeTmp'

  Run Build Command(s): /usr/local/bin/cmake -E env VERBOSE=1 /usr/bin/make
  -f Makefile cmTC_a6e2f/fast

  make[1]: Entering directory
  '/home/yi/Downloads/example-app/build/CMakeFiles/CMakeTmp'

  /usr/bin/make -f CMakeFiles/cmTC_a6e2f.dir/build.make
  CMakeFiles/cmTC_a6e2f.dir/build

  make[2]: Entering directory
  '/home/yi/Downloads/example-app/build/CMakeFiles/CMakeTmp'

  Building CXX object CMakeFiles/cmTC_a6e2f.dir/detect_cuda_version.cc.o

  /usr/bin/c++ -I/usr/local/cuda-11.6/include -o
  CMakeFiles/cmTC_a6e2f.dir/detect_cuda_version.cc.o -c
  /home/yi/Downloads/example-app/build/detect_cuda_version.cc

  Assembler messages:

  Fatal error: can't create
  CMakeFiles/cmTC_a6e2f.dir/detect_cuda_version.cc.o: No such file or
  directory

  make[2]: *** [CMakeFiles/cmTC_a6e2f.dir/build.make:78:
  CMakeFiles/cmTC_a6e2f.dir/detect_cuda_version.cc.o] Error 1

  make[2]: Leaving directory
  '/home/yi/Downloads/example-app/build/CMakeFiles/CMakeTmp'

  make[1]: *** [Makefile:127: cmTC_a6e2f/fast] Error 2

  make[1]: Leaving directory
  '/home/yi/Downloads/example-app/build/CMakeFiles/CMakeTmp'

Call Stack (most recent call first):
  /usr/local/libtorch/share/cmake/Caffe2/Caffe2Config.cmake:88 (include)
  /usr/local/libtorch/share/cmake/Torch/TorchConfig.cmake:68 (find_package)
  CMakeLists.txt:4 (find_package)


-- Configuring incomplete, errors occurred!
make: *** [Makefile:179: cmake_check_build_system] Error 1

 *  The terminal process "/usr/bin/bash '-i', '-c', 'cmake --build . --config Debug'" terminated with exit code: 2.

  6. I tried g++ command, but I had problem with linking torch libraries (-L) and including header files (-I).

    1
2
3
4
5
6
7

     *  Executing task: C/C++: g++ build active file 

Starting build...
/usr/bin/g++ -fdiagnostics-color=always -g /home/yi/Downloads/example-app/main.cpp -o /home/yi/Downloads/example-app/main
/home/yi/Downloads/example-app/main.cpp:1:10: fatal error: torch/torch.h: No such file or directory
    1 | #include <torch/torch.h>
      |          ^~~-~~-~~-~~-~~-~~

Ref:

  1. CMake+VSCode编译运行C++程序简单教程 - 迷楼的文章 - 知乎

  2. Linux环境下使用VScode调试CMake工程 - Colorful的文章 - 知乎

  3. libtorch 常用api函数示例（史上最全、最详细）- 无左无右 - 博客园

Inspect Tensor

(2023-11-19)

  1. Save tensors to a file.

    (LibTorch -> PyTorch tensors) How to save a C++ Libtorch Tensor and load it into a Python Pytorch project?

  2. In GDB: *(float_t*)x.data_ptr()

    1
2

    int main(){
const torch::Tensor bkg = torch::full({3}, 0., torch::device(torch::kCUDA));

    GDB: *(float_t*)bkg.data_ptr() return 0.

    Is it possible to view values of a at::Tensor in Visual codes debug variable view (linux)?

© 2020 - 2025 Zichen Wang
Built with Hugo
Theme Stack designed by Jimmy