Build LibTorch for JIT

Debug Torch Script(Windows)

Prerequisites

• CMake >= 3.14

• Python >= 3.6

Step 1: Clone the code

git clone --recursive https://github.com/pytorch/pytorch.git
cd pytorch

Step 2: Generate Visual Studio Solution

Open a command prompt, and run the following commands to open cmake-gui.

REM cwd=pytorch root folder
set BUILD_TEST=OFF
set USE_DISTRIBUTED=OFF
set USE_CUDA=OFF
set USE_OPENMP=OFF
set USE_MKLDNN=OFF
set USE_FBGEMM=OFF
set USE_QNNPACK=OFF
set BUILD_PYTHON=OFF
set BUILD_CAFFE2_OPS=OFF

REM set USE_NNPACK=OFF
REM set USE_XNNPACK=OFF

set USE_NINJA=OFF
set CMAKE_GENERATOR="Visual Studio 16 2019"

python setup.py install --cmake-only
cmake-gui

Uncheck BUILD_PYTHON, and then run Configure->Generate->Open Project to open the Visual Studio solution.

Step 3: Build libtorch with Debug configuration

Step 4: Find one unittest to debug

For example, use test/cpp/api/jit.cpp as an entrance to start with.

LibTorch Minimal Build

Windows

cmake -DBUILD_SHARED_LIBS:BOOL=ON `
    -DBUILD_TEST=OFF `
    -DUSE_DISTRIBUTED=OFF `
    -DUSE_CUDA=OFF `
    -DUSE_OPENMP=OFF `
    -DUSE_MKLDNN=OFF `
    -DUSE_FBGEMM=OFF `
    -DUSE_QNNPACK=OFF `
    -DBUILD_PYTHON=OFF `
    -DUSE_XNNPACK=OFF `
    -DUSE_NNPACK=OFF `
    -DBUILD_CAFFE2=OFF `
    -DPYTHON_EXECUTABLE:PATH=C:\Path\to\Python\python.exe `
    -DCMAKE_INSTALL_PREFIX:PATH=libtorch_install `
    -G "Visual Studio 16 2019" -A x64 `
    -S . -B libtorch_build
cmake --build libtorch_build --config Release --target install

A functional cmake file with the above configurations is prepared for you to build libtorch directly.

cmake -S ./third_party/libtorch/ -B libtorch_build `
    -DCMAKE_INSTALL_PREFIX=libtorch_install `
    -G "Visual Studio 16 2019" -A x64
cmake --build libtorch_build