Best Practices to Edit and Compile PyTorch Source Code On Windows

Page Maintainers:: @mszhanyi , @ozanMSFT

• Goals
• Prerequisites
  • (Optional) CUDA
  • (Optional) Other Tools
  • MSVC - Build Tools
  • Ninja
  • Python
  • Source Code
  • Installing requirements.txt
• Visual Studio 2022/2019 Guide
  • Open Project
  • Build Project
  • Modify Source and Rebuild
  • Run Project
• Visual Studio Code Guide
  • Open Project
  • Build Project
  • Modify Source and Rebuild
  • Run Project
• Final Notes

Goals

In this guide, we'll cover some best practices to work with PyTorch C++ Source code on Windows.

Note: In this guide, we'll only cover libtorch compilation as an example. As a general base guide; you can also refer the page building libtorch.

Prerequisites

(Optional) CUDA

Installing CUDA is not mandatory for compilation, but you may want to install it depending on your needs.

Please take a look on this page for the supported CUDA versions and installation: https://github.com/pytorch/pytorch#from-source

(Optional) Other Tools

Although MKL, cuDNN, oneAPI, ONNX Runtime and some other dependencies may not be necessary for this guide's scope, you may want to install them depending on your needs for other compilation scenarios. [Note: MKL installation is recommended]

MSVC - Build Tools

You need to have Microsoft Visual C++ (MSVC) compiler toolset on your system.

With Visual Studio Installer, you need to install Desktop Development with C++ workload.

Option 1: If you already have Visual Studio 2022 or Visual Studio 2019

You can open Visual Studio Installer by searching from Start Menu and click Modify for your relevant Visual Studio installment.

Then you can check if Desktop Development with C++ workload is checked.

Option 2: If you're planning to use Visual Studio Code

You can install Build Tools for Visual Studio from following URL without installing full Visual Studio IDE.

https://visualstudio.microsoft.com/downloads/#build-tools-for-visual-studio-2022

After installation, you can open Visual Studio Installer and install Desktop Development with C++ workload.

Note: Please check installation path, we'll need it when defining PATH.

Ninja

Option 1: Use Ninja executable from Desktop Development with C++ workload

If you followed MSVC - Build Tools, Ninja should be already under build tools installation folder.

If you've Visual Studio 2022 or Visual Studio 2019: Ninja should be already under your Visual Studio folder:

{VSPATH}\Common7\IDE\CommonExtensions\Microsoft\CMake\Ninja\ninja.exe

If you've Visual Studio Code: Ninja should be already under installation folder which you've specified while installing Desktop Development with C++ workload.

{BUILDTOOLSPATH\Common7\IDE\CommonExtensions\Microsoft\CMake\Ninja\ninja.exe

Option 2: Use Ninja from GitHub Release

Download Ninja separately and add it to PATH.

https://github.com/ninja-build/ninja/releases

Python

You need to install Python to your system.

https://www.python.org/downloads/windows/

Note: This guide is tested with the latest stable Python version "3.12" (November 2023)

Source Code

Source code should be fetched as instructed in the main page. Get the PyTorch Source

Installing requirements.txt

Install requirements.txt from root folder.

pip install -r requirements.txt

Visual Studio 2022/2019 Guide

Open Project

You can open PyTorch project with Open a local folder option.

Build Project

After the opening source code folder, Visual Studio should automatically detect CMake type project.

1. Click Manage Configurations... to open CMake Settings window.

   

2. Click Edit JSON shortcut to edit parameters of CMake.

   

3. Copy and paste following example JSON content into CMakeSettings.json file and save it after replacing placeholders accordingly. Especially, please do not forget to replace {{PYTHON_ROOT}} parameter as suggested below.

   Note: Please replace {{PYTHON_ROOT}} with actual root folder of Python. You can find it with cmd command where python

   Note: Settings and parameters may change depending on your needs. Currently, we're focusing for an example libtorch build.

   Note: For this example, following parameters are different than default parameters installRoot, cmakeCommandArgs and variables section

   {
      "configurations": [
         {
            "name": "x64-Debug",
            "generator": "Ninja",
            "configurationType": "Debug",
            "inheritEnvironments": [ "msvc_x64_x64" ],
            "buildRoot": "${projectDir}\\out\\build\\${name}",
            "installRoot": "${projectDir}\\torch",
            "cmakeCommandArgs": "-DVS_STARTUP_PROJECT=Torch",
            "ctestCommandArgs": "",
            "variables": [
               {
                  "name": "BUILD_PYTHON",
                  "value": "False",
                  "type": "BOOL"
               },
               {
                  "name": "BUILD_TEST",
                  "value": "True",
                  "type": "BOOL"
               },
               {
                  "name": "CMAKE_PREFIX_PATH",
                  "value": "{{PYTHON_ROOT}}\\Lib\\site-packages",
                  "type": "PATH"
               },
               {
                  "name": "NUMPY_INCLUDE_DIR",
                  "value": "{{PYTHON_ROOT}}\\Lib\\site-packages\\numpy\\core\\include",
                  "type": "PATH"
               },
               {
                  "name": "PYTHON_EXECUTABLE",
                  "value": "{{PYTHON_ROOT}}\\python.exe",
                  "type": "PATH"
               },
               {
                  "name": "PYTHON_INCLUDE_DIR",
                  "value": "{{PYTHON_ROOT}}\\Include",
                  "type": "PATH"
               },
               {
                  "name": "USE_NUMPY",
                  "value": "True",
                  "type": "BOOL"
               }
            ]
         }
      ]
   }
   

4. After saving the file, CMake output should automatically start to prepare configuration, wait for a while and you should see CMake generation finished message on the bottom.

   Note: If you're seeing an error, please check parameters before continue.

   

5. Now, navigate Build -> Build All menu to start build. Build result will be under out\build\x64-Debug folder.

   Note: First time build might take some time depending on build machine's specifications.

   

Modify Source and Rebuild

1. As an example change; let's open test\cpp\api\nn_utils.cpp file and search for TEST_F(NNUtilsTest, ConvertParameters) test. Let's add a printf in this test.

   

2. Build project again by navigating Build -> Build All.

   Note: Following builds after the first time build should take only a few minutes.

   

Run Project

1. Let's setup Debug Target as test_api.exe (bin\test_api.exe) for our project's scope.

   Note: You can hide all unnecessary targets for clean view.

   

   

2. Let's also modify Debug and Launch Settings for test_api to specifiy additional parameters for debugging, this will open launch.vs.json file.

   

3. For simplicity, let's add this filter as an argument to run only our modified test and save launch.vs.json file.

   "args": [ "--gtest_filter=NNUtilsTest.ConvertParameters" ]
   

   

4. Run project and see the result.

   

   

Visual Studio Code Guide

Open Project

1. Open PyTorch folder with Visual Studio Code.

2. Install CMake Tools extension by Microsoft.

Build Project

1. Navigate .vscode folder from root and create settings.json file. Copy following JSON content and save it after replacing placeholders accordingly. Especially, please do not forget to replace {{PYTHON_ROOT}} parameter as suggested below.

   

   Note: Please replace {{PYTHON_ROOT}} with actual root folder of Python. You can find it with cmd command where python

   Note: Settings and parameters may change depending on your needs. Currently, we're focusing for an example libtorch build.

   {
      "cmake.generator": "Ninja",
      "cmake.configureSettings":{
         "BUILD_PYTHON":"False",
         "BUILD_TEST":"True",
         "CMAKE_PREFIX_PATH":"{{PYTHON_ROOT}}\\Lib\\site-packages",
         "NUMPY_INCLUDE_DIR":"{{PYTHON_ROOT}}\\Lib\\site-packages\\numpy\\core\\include",
         "PYTHON_EXECUTABLE":"{{PYTHON_ROOT}}\\python.exe",
         "PYTHON_INCLUDE_DIR":"{{PYTHON_ROOT}}\\Include",
         "USE_NUMPY":"True"
      }
   }
   

2. In the Status Bar at bottom, select Debug as build variant and desired toolkit based on MSVC - Build Tools setup.

   

3. Open Command Palette with Ctrl + Shift + P and search for CMake: Configure and run it. Please wait until CMake configure is done successfully by following up output at the bottom.

   

   

4. After the configure step is done; Click Build button in the Status Bar. Build result will be under build folder.

   Note: First time build might take some time depending on build machine's specifications.

   

Modify Source and Rebuild

1. As an example change; let's open test\cpp\api\nn_utils.cpp file. Then, search for TEST_F(NNUtilsTest, ConvertParameters) test. Let's add a printf in this test.

   Note: You can use Ctrl + P and type nn_utils.cpp as a shortcut.

   

2. Build project again by clicking Build in the Status Bar.

   Note: Following builds after the first time build should take only a few minutes

   

Run Project

1. Open settings.json to configure Debug Config. For this example, let's add --gtest_filter=NNUtilsTest.ConvertParameters filter to the cmake.debugConfig.

   "cmake.debugConfig": {
      "args": ["--gtest_filter=NNUtilsTest.ConvertParameters"]
   }
   

   

2. Click Launch the selected target button in the Status Bar and search test_api to select test_api.exe.

   

   

3. Previous step will automatically trigger run, you see result at Debug Console.

   Note: You can run again Ctrl + F5 shortcut or with Launch the selected target button any time.

   

Final Notes

You might find some changes you made do not work and after recompiling, the changes dismissed. That is because some PyTorch C++ code is generated during compilation. For more information on code generation, you could check out Codegen and Structured Kernels