Patreon
Home / Guides / AI Requirements Installer for Windows and Ubuntu: Python, CU

AI Requirements Installer for Windows and Ubuntu: Python, CUDA and C++ in One Run

By Furkan Gözükara · Updated Nov 21, 2026

The SECourses AI Requirements Installer prepares a Windows or Ubuntu PC for local AI apps in one run: it installs Python 3.10 to 3.14, CUDA 12.9 and 13.2, cuDNN 9.20, the C++ compilers, Git, Node.js, .NET and FFmpeg, sets every path and variable, and tests the result. On Windows you run one .bat file as administrator; on Ubuntu you run one sudo command. The installer is available to members on Patreon, and the full installer tutorial with its video is public.

What it installs on each system

ComponentWindowsUbuntu
Supported systemsWindows 10 (21H2 or newer) and 11; 11 recommendedUbuntu 22.04, 24.04 and 26.04 LTS (24.04 recommended): desktop, server, cloud GPU machines, WSL2, Linux Mint, Zorin OS
Python3.10 to 3.14 for all users in C:\Python310 to C:\Python314; 3.12 is the default3.10 to 3.14 plus free-threaded 3.13t and 3.14t; 3.12 is the default and Ubuntu's own python3 stays as it is
CUDA12.9 and 13.2 side by side, 13.2 the default12.9 and 13.2 (13.4 on Ubuntu 26.04); use-cuda 12.9 switches one terminal
cuDNN9.20, the only cuDNN registered9.20 for CUDA 13
C++ toolsVisual Studio 2022 Build Tools (MSVC, Windows SDK), latest Visual C++ runtimeGCC, Clang, CMake and Ninja
NVIDIA driverNewest compatible WHQL driver for GeForce RTX 20 series and newerNewest driver Ubuntu offers, R595 or newer (R615 on 26.04), with open kernel modules
AlsoGit, Node.js 24, .NET 10 SDK, FFmpeg 9 shared and static, long paths, a 100 GB page fileGit with Git LFS, Node.js 24, .NET 10 SDK, uv, FFmpeg 9, swap up to 100 GB in total

The installers are made for the SECourses apps and their full features, such as ComfyUI, SwarmUI, SECourses Musubi Trainer, AI Toolkit and Ultimate TTS, with builds of torch.compile, Triton, flash-attention, sage-attention, xformers and insightface in mind. Without an NVIDIA GPU everything still installs and compiling works, but CUDA apps need an NVIDIA GPU.

How to run it on Windows

  1. Download the latest zip (Requirements_Installers_v4.zip) from the AI Requirements Installers post. Put it in your Downloads folder, not directly on C:, and extract it.
  2. Right-click Windows_AI_Requirements_Installer.bat and choose Run as administrator. Keep Windows_AI_Requirements_Installer.py in the same folder. A normal double-click also works: it asks for administrator rights itself. The .bat brings a temporary portable Python, so nothing has to be installed first.
  3. Wait for the 20 steps to finish. The HTML report then opens in your browser, and logs.txt is saved next to the .bat.
  4. Open a new Command Prompt or PowerShell window to use the new PATH, and restart Windows when the report asks. If a new NVIDIA driver required the restart, run the installer again afterwards to finish the tests.

This is the only SECourses app that needs administrator rights; the other apps start with a double-click. On a Windows 11 PC with an RTX 3090 Ti, the whole install, update and test run took 20 minutes 45 seconds, with the CUDA and cuDNN installers already downloaded by an earlier run. CUDA 12.9 and 13.2 took 5 minutes, the five Python versions 5 minutes and the tests 1 minute 48 seconds.

How to run it on Ubuntu

Put Ubuntu_AI_Requirements_Installer.sh into a folder, open a terminal there and run:

sudo bash ./Ubuntu_AI_Requirements_Installer.sh

Enter your password and the rest is automatic: 17 steps, with the NVIDIA driver last. On a fresh PC, the seven Python versions took 43 seconds and the tests 28 seconds. The installation keeps running even if you close the terminal. Restart Ubuntu when the report asks for it after a new driver.

The environment is written once for all users and loaded by login shells, bash and zsh. The libraries are registered system wide, so no LD_LIBRARY_PATH is needed, and lines in ~/.bashrc and similar files that switch to another CUDA version are removed with a backup.

PATH order, variables and existing tools

  • PATH order: cuDNN and CUDA 13.2 come first, then the Windows system folders, then Python 3.12 and the other versions, the C++ build paths and the FFmpeg shared build. Your own entries stay after these, and the report colors added, moved and kept entries.
  • Variables: CUDA_HOME points to CUDA 13.2 and CUDNN_PATH to the cuDNN 9.20 root. INCLUDE, LIB and LIBPATH make the C++ compiler work in every new terminal, and VSLANG keeps compiler messages in English for torch.compile.
  • Visual Studio updates: fixed paths and a scheduled task keep compiling working after Visual Studio installs a new compiler version.
  • Existing tools: an existing Visual Studio is updated, a current Visual C++ runtime is left alone, and an older Python from the Python Install Manager is backed up while its virtual environments keep working.
  • One cuDNN: only cuDNN 9.20 for CUDA 13 is registered, so PyTorch always loads one matching cuDNN.

78 Windows tests and 66 Ubuntu tests

The installer tests everything with exactly the PATH and variables a new terminal gets. On Windows, all 78 tests passed on the test PC:

  • nvcc finds CUDA 13.2, cl is the MSVC compiler and ffmpeg is the shared build.
  • Every Python version runs with ssl, sqlite3, ctypes and tkinter, creates a virtual environment and runs pip in it.
  • A CUDA program is compiled and run with both CUDA versions, with correct results on the RTX 3090 Ti GPU, and GPU code for RTX 50 series (Blackwell) cards compiles too.
  • An OpenMP library builds the way torch.compile uses it, a Python C extension builds the way pip builds insightface, and CUDA extensions build the way flash-attention and xformers are built.
  • FFmpeg encodes a test video, a .NET app builds and runs as SwarmUI needs, and a path longer than 260 characters works.

On Ubuntu, all 66 tests passed, including a compiled C extension for every Python version and the use-cuda 12.9 switch. On Windows, the optional --torch-test installs PyTorch 2.13 with CUDA 13 in a temporary environment and runs real cuDNN convolution and BLAS tests on your GPU.

Useful Windows options

  • --dry-run shows the plan and changes nothing.
  • --skip nvidia keeps your driver; --only nvidia updates just the driver.
  • --only env,verify runs only the named steps, --skip skips steps and --list-steps lists them.
  • --restore-env "backup folder" puts PATH and the variables back.

Since V2, Windows_AI_Requirements_Installer_Skip_Driver_Update_Virtual_RAM.bat skips the driver and page file steps, and you can edit it to add more skips.

Safety, reruns and recent updates

Every installer must carry a valid digital signature from its publisher, archives pass an integrity check, and every change is backed up first with a HOW_TO_RESTORE.txt in the backup folder. You can run the installer again at any time: finished parts are only checked, failed or interrupted parts are repaired, and components and the driver are updated.

  • V3 (25 October 2026): downloads prefer IPv4. In tests on one connection, PyPI downloads ran at 20 KB/s over IPv6 and 1.3 MB/s over IPv4, so IPv6 was about 64 times slower. IPv6 stays enabled, and --skip ipv4 leaves this out.
  • V4 (6 November 2026): a new uvsettings step clears UV_SYSTEM_PYTHON and UV_PYTHON and turns off the [pip] lines system = true, python, target and prefix in uv.toml and pyproject.toml, keeping a .bak copy. It repairs uv and pip setups that other installers broke.

The post says the installer will be updated as needed, for example for Torch 2.14 and CUDA 13.4.

FAQ

Do I need to run the requirements installer for every AI app?

No. It is a one-time setup for the PC, not for each app. Run a newer version when the post shows one, and it updates what changed.

Can I keep my current NVIDIA driver?

Yes. Use --skip nvidia, or the Skip_Driver_Update_Virtual_RAM bat file, which also skips the page file. The installer never downgrades a working driver.

Does it work on cloud GPUs and WSL2?

Yes. The Ubuntu installer runs on cloud GPU machines and WSL2 as well as on desktops and servers, on 64-bit x86.

What should I do if a step fails?

Run the installer again; failed parts are repaired. If it still fails, send logs.txt with your question on the SECourses Discord.

Can I still set things up by hand?

Yes. The manual route with Visual Studio, CUDA, cuDNN and PATH editing is in the CUDA and C++ build errors guide, and the older manual setup video shows it step by step.

Related