1. Prerequisites
2. Cloning PyTorch
3. Preparing Environment
4. Setting Environment Variables
5. Build
   1. PyTorch
   2. LibTorch

1. Build Tools for Visual Studio 2022OR any version ofVisual Studio 2022

   • UnderWorkloads, selectDesktop development with C++

   • UnderIndividual Components, selectARM64/ARM64EC build tools (latest)

2. Python 3.12 Windows installer (ARM64)

3. Download and installArm Performance Libraries for Windows

   • %ARMPL_DIR% environment variable should be automatically added to yourPATH by default installation settings.

4. InstallRust

5. Buildinglibuv (v1.39.0) from source (we'll use the folder path inBuild)

   • OpenARM64 Native Tools Command Prompt for VS2022 from start menu
   • Change path to the suitable folder
   • Run following command to buildlibuv

   git clone https://github.com/libuv/libuv.git -b v1.39.0cd libuvecho Clean folder...git clean -fxdecho Configuring libuv...mkdir libuv\buildcd libuv\buildcmake .. -DBUILD_TESTING=OFFecho Building libuv...cmake --build . --config Releaseecho Installing libuv...cmake --install . --prefix ../installpause

6. (OPTIONAL) You may installsccache for speeding up future builds by cache support.

   • You need to addsccache to yourPATHmanually.

1. Clone the PyTorch repository (or your fork)

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

2. Init and sync submodules

   git submodule syncgit submodule update --init --recursive

NOTE: Since currentlycondadoesn't support Windows ARM64, instead we're recommending to use built-in Pythonvenv functionality for similar isolating experience. (more info)

1. Creating a virtual environment (venv)

   python -m pip install --upgrade pippython -m venv .venvecho *> .venv\.gitignorecall .\.venv\Scripts\activate

2. Install requirements

   pip install -r requirements.txt

3. ActivateArm64 Native Tools Command Prompt for VS 2022 environment by calling one of the following commands based on what you've installed.

   • For Build Tools:

     "C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64

   • For Visual Studio 2022 Community:

     "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\Build\vcvarsall.bat" arm64

   • For Visual Studio 2022 Enterprise:

     "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" arm64

   • For Visual Studio 2022 Preview:

     "C:\Program Files\Microsoft Visual Studio\2022\Preview\VC\Auxiliary\Build\vcvarsarm64.bat" arm64

1. Set the following environment variables:

   setBLAS=APLsetUSE_LAPACK=1setMSSdk=1setDISTUTILS_USE_SDK=1:: please change {YOUR_LIBUV_FOLDER} to the your libuv foldersetlibuv_ROOT={YOUR_LIBUV_FOLDER}\install

2. Copylibuv dll to thelib folder to include it for packaging (please check:Prerequisites)

   copy%libuv_ROOT%\lib\Release\uv.dll torch\lib\uv.dll

3. (OPTIONAL) If you want to usesccache for speeding up your future build by cache also set following.

   setCMAKE_C_COMPILER_LAUNCHER=sccachesetCMAKE_CXX_COMPILER_LAUNCHER=sccache

You can generate a.whl (wheel) file forPyTorch (the Python package)OR a.zip file forLibTorch (the C++ core). If you'd like to use PyTorch right away without creating a wheel file, you can build it directly in your environment. Here are your options:

1. OPTION 1 - Generating and Installing Wheel File (for distribution or installation on other systems):

   Call following command to start build and generate output

   :: set PYTORCH_BUILD_VERSION=2.6.0   ## OPTIONAL: If you wish to give a specific versioning for `.whl` file :: set PYTORCH_BUILD_NUMBER=1        ## (for instance you can specify 2.6.0 as shown) python setup.py bdist_wheel

   Output file will be underdist folder with.whl extension.

   Format will look liketorch-FOO+gitBAR-cp312-cp312-win_arm64.whlOR if you specify version, format will be liketorch-2.6.0-cp312-cp312-win_arm64.whl

   Now you can copy and install this.whl file anywhere in your system by calling

   pip install torch-YOURVERSION-cp312-cp312-win_arm64.whl

2. OPTION 2 - Directly Installing to Local Environment (for development on your system only)

   Run one of the following commands to initiate the build:

   • Install the package in development mode, ideal when making frequent changes to codebase

   python setup.py develop

   • Install the package for regular use, rather than ongoing development

   python setup.py install

1. Additional preparation forlibtorch folder

   :: Options: Release, Debug or RelWithDebInfosetCMAKE_BUILD_TYPE=Release:: Prepare the environmentmkdir libtorchmkdir libtorch\binmkdir libtorch\cmakemkdir libtorch\includemkdir libtorch\libmkdir libtorch\sharemkdir libtorch\test

2. Call following command to start build and generate output

   python ./tools/build_libtorch.py

3. Prepare.zip file fromlibtorch output

   :: Moving the files to the correct locationmove /Y torch\bin\*.* libtorch\bin\move /Y torch\cmake\*.* libtorch\cmake\robocopy /move /e torch\include\ libtorch\include\move /Y torch\lib\*.* libtorch\lib\robocopy /move /e torch\share\ libtorch\share\move /Y torch\test\*.* libtorch\test\move /Y libtorch\bin\*.dll libtorch\lib\:: Create output under distmkdir disttar -cvaf dist/libtorch-win.zip -C libtorch *:: Cleanup raw data to save spacermdir /s /q libtorch

4. Now you can copy and extractdist/libtorch-win.zip file anywhere in your system

   • For C++ example, you can visitInstalling C++ Distributions of PyTorch

   • You can just add following line to yourCMakeLists.txt file.

     set(CMAKE_PREFIX_PATH"C:/YOUR_CORRESPONDING_PATH/libtorch-win")