One-Click AI PC Setup: Hands-On with the Windows and Ubuntu AI Environment Installers

Title

One-Click AI PC Setup: Hands-On with the Windows and Ubuntu AI Environment Installers

Meta Description

SECourses releases one-click AI environment installers for Windows and Ubuntu, automatically configuring CUDA 12.9/13.2, cuDNN, Python 3.10-3.14, and the C++ toolchain, with availability verified through 78 and 66 real-world tests. This article details the installation process, PATH tuning logic, and measured results.

Article Content

Running AI applications locally is often most painful at the environment setup stage: CUDA, cuDNN, multiple Python versions, and the C++ compiler toolchain all intertwine, and one wrong move plunges you into dependency conflicts. To address this pain point, SECourses has released a one-click AI environment installer—double-click to run on Windows, a single command on Ubuntu—that claims to turn a new PC directly into an "AI-ready" state. Based on its official tutorial and hands-on test report, this article breaks down the design approach and actual performance of this toolset.

Why a Dedicated AI Environment Installer Is Needed

Manually setting up an AI development environment involves dozens of components: NVIDIA drivers, the CUDA toolkit, cuDNN, Python interpreters, Visual Studio Build Tools, Git, Node.js, FFmpeg, and more. Each component has version compatibility requirements, and an incorrect PATH order can cause PyTorch to load the wrong cuDNN version.

According to the author, his AI application matrix (ComfyUI, SwarmUI, Musubi Trainer, Ultimate TTS, etc.) relies on components that are sensitive to the build environment, such as torch.compile, Triton, flash-attention, sage-attention, xformers, and insightface. Traditional manual installation tutorials (the predecessor of this installer) had too high a barrier to entry, which prompted the author to script the entire process.

The core value of this installer lies in "configure once, benefit all applications." It is not aimed at a single application but instead establishes a unified AI runtime foundation for the entire PC, so that all subsequent local AI applications reuse the same environment.

The version coupling between CUDA, cuDNN, and PyTorch

CUDA is NVIDIA's parallel computing platform, and cuDNN is the deep neural network acceleration library built on top of it. Frameworks such as PyTorch bind to specific CUDA and cuDNN versions at compile time; at runtime, if an mismatched cuDNN dynamic library (such as cudnn64_9.dll) is loaded, the result ranges from degraded performance to a direct crash. The installer places cuDNN 9.20 at the front of PATH and registers only that version, precisely to prevent multiple cuDNN copies from being loaded unexpectedly.

Compilation dependencies of torch.compile and Triton

torch.compile is the just-in-time compilation feature introduced in PyTorch 2.0. It relies on Triton (a GPU kernel compiler) and a complete C++/CUDA toolchain to generate optimized kernels. Libraries such as flash-attention and xformers likewise need to compile CUDA extensions at install time. Therefore, a usable AI environment requires not only runtime libraries but also MSVC (Windows) or GCC/Clang (Linux) plus the nvcc compiler from the CUDA Toolkit.

Windows Installer: 20 Steps of Automation and 78 Real Tests

How to Run and What Gets Installed

Users simply extract the zip package into the Downloads folder (do not place it directly in the root of the C drive), then right-click Windows_AI_Requirements_Installer.bat and select "Run as administrator." The script ships with a portable Python, so no dependencies need to be installed beforehand.

The installer executes 20 steps, covering: the latest compatible NVIDIA WHQL driver (RTX 20 series and above), Python 3.10 through 3.14 (3.12 is the default), CUDA 12.9 and 13.2 installed side by side (13.2 is the default), cuDNN 9.20 (only this version is registered to avoid PyTorch loading conflicts), Visual Studio 2022 Build Tools (MSVC and the Windows SDK), the latest VC++ runtime, Git, Node.js 24, .NET 10 SDK, FFmpeg 9 (both shared and static library versions), PATH order tuning, Windows long path support, and a 100GB page file.

One detail worth mentioning is the PATH handling logic: the installer places cuDNN and CUDA 13.2 first, followed by the Windows system directories, Python 3.12 and other versions, the C++ build paths, and the FFmpeg shared libraries. The user's existing entries for other programs are retained afterward and are not deleted. The pre-installation PATH backup is fully listed in the report and can be restored with a single --restore-env command.

Real Tests, Not Just Version Checks

The testing phase is what distinguishes this installer from ordinary install scripts. All 78 tests run in "the environment actually obtained by a new terminal," not self-tested in the environment at script startup. The tests include: compiling and launching a CUDA program (verifying GPU computation results are correct on an RTX 3090 Ti), compiling Thrust and CUB code, compiling GPU code targeting the RTX 50 series (Blackwell), compiling kernels at runtime with NVRTC, building a Python C extension (simulating the pip build process for insightface), building CUDA extensions (simulating how flash-attention and xformers are built), encoding a test video with FFmpeg and reading it back, creating and running a .NET console application, and verifying that long paths exceeding 260 characters work.

Test results are presented item by item in an HTML report, and a logs.txt is also generated. According to the author's hands-on test, a Windows 11 PC equipped with an RTX 3090 Ti took 20 minutes and 45 seconds to complete installation, updates, and testing (the CUDA and cuDNN installers had already been downloaded in advance by a previous run).

NVRTC and runtime compilation

NVRTC (NVIDIA Runtime Compilation) allows a program to compile CUDA C++ source code into GPU-executable code at runtime, without pre-compiling offline with nvcc. Many AI libraries use NVRTC to implement JIT kernel generation. The test item "compile kernels at runtime with NVRTC" verifies whether the CUDA environment configured by the installer can support this kind of dynamic compilation path, which is critical for applications that rely on JIT technologies such as Triton.

Thrust and CUB

Thrust is CUDA's C++ template library, providing parallel algorithm interfaces similar to the STL; CUB is its underlying GPU primitive library. Both are header-only libraries, but compiling with them still requires a correct CUDA toolchain and C++ standard support. Testing the compilation of Thrust and CUB code is essentially checking whether the template instantiation and device code generation pipeline is complete.

Design Details for Compatibility with Existing Environments

For PCs that already have development tools installed, the installer adopts an "update rather than overwrite" strategy: existing Visual Studio Build Tools are updated to the latest 17.x version, an already up-to-date VC++ runtime is left untouched, an old Python 3.14 managed by Python Install Manager is backed up and unregistered, while a link is created to ensure the old Python's virtual environments remain usable. The .NET SDK is installed side by side rather than replaced.

Visual Studio compiler updates are installed to a new version folder, which invalidates paths in PATH. To address this, the installer sets up a fixed path pointing to the current compiler and uses a scheduled task to re-point it to the new version at startup, at login, after VS updates, and daily, ensuring that compilation-dependent features such as torch.compile remain continuously available.

Ubuntu Installer: One Command and 66 Tests

Supported Range and How to Run

The Ubuntu version supports 22.04, 24.04, and 26.04 LTS, covering desktop, server, cloud GPU instances, WSL2, and Ubuntu-based Linux Mint and Zorin OS (64-bit x86). In a terminal, navigate to the installer's directory and run sudo bash ./Ubuntu_AI_Requirements_Installer.sh. Closing the terminal will not interrupt the installation process.

The installation includes Python 3.10 through 3.14 plus the free-threaded builds 3.13t and 3.14t (Ubuntu's built-in python3 remains unchanged), CUDA 12.9 and 13.2 (13.4 on Ubuntu 26.04), cuDNN 9.20, GCC, Clang, CMake, Ninja, Git and Git LFS, Node.js 24, .NET 10 SDK, FFmpeg 9, uv, and a 100GB swap file. The NVIDIA driver installs the latest version provided by Ubuntu (R595 or newer, R615 on 26.04), using Ubuntu-signed open-source kernel modules; each run only upgrades and never downgrades, and it is executed as the final step.

Environment Configuration and Testing

Environment variables are written to profile files shared by all users and are loaded by login shells, bash, and zsh. CUDA_HOME, CUDA_PATH, CUDNN_PATH, and the CUDA tool paths take effect for all users. Library files are registered at the system level, so there is no need to set LD_LIBRARY_PATH. The use-cuda 12.9 command can temporarily switch a single terminal to CUDA 12.9. Old configuration lines in the user's ~/.bashrc that switch CUDA versions are backed up and then removed.

All 66 tests run as an ordinary user in the environment actually obtained by a new terminal, covering the ssl, sqlite3, lzma, readline, and tkinter modules of each Python version, virtual environment creation and C extension compilation, compilation and launch of both CUDA versions, compilation of RTX 50 series code, verification of use-cuda switching, C++17 compilation with GCC and Clang, building and loading OpenMP libraries, building a Triton-style GPU launcher, availability of FFmpeg encoders (x264, x265, SVT-AV1, VP9, NVENC), and functional verification of tools such as Git LFS, Node.js, .NET, aria2c, 7-Zip, unrar, jq, cloudflared, nvtop, and htop.

Free-threaded Python (3.13t/3.14t)

Free-threaded Python is an experimental build of CPython that removes the Global Interpreter Lock (GIL), allowing multiple threads to truly execute Python bytecode in parallel. This has potential benefits for scenarios such as data preprocessing and multithreaded inference scheduling, but many C extensions have not yet been adapted. The installer installs 3.13t and 3.14t side by side, giving users the option to experiment with GIL-free concurrency in an isolated environment without affecting the default 3.12 environment.

The use-cuda switching mechanism

The Ubuntu installer deploys both CUDA 12.9 and 13.2, but by default PATH points to 13.2. The use-cuda 12.9 command temporarily points the CUDA tool path to version 12.9 by modifying the current shell's environment variables, allowing a specific application to choose the older CUDA version without affecting the system default. This solves compatibility issues when certain frameworks do not yet support the latest CUDA version.

Safety and Recoverability

All downloads on both platforms are verified: Windows checks digital signatures (from publishers such as NVIDIA, Microsoft, and the Python Software Foundation) and archive integrity; Ubuntu uses SHA-256 checksums. Backups are made before all changes, and the backup directory includes HOW_TO_RESTORE.txt and supports --restore-env to restore environment variables and files. Ubuntu also provides --uninstall to remove content the installer added outside of APT, and --repair to check and reinstall corrupted packages after a power outage or crash.

Common Options and Re-run Strategy

The Windows side provides options such as --dry-run (show the plan only), --only nvidia (update the driver only), --skip nvidia (keep the current driver), --torch-test (install PyTorch 2.13 in a temporary environment and run real cuDNN convolution and BLAS tests), --restore-env (restore backups), and --no-browser (do not automatically open the report). The Ubuntu side has corresponding options: --yes (unattended), --dry-run, --skip/--only, --swap-gb (custom swap file size), --torch-test, --repair, --restore-env, and --uninstall.

Both installers are designed to be run repeatedly: completed parts are only checked, failed or interrupted parts are retried and repaired, components are updated to the versions built into the installer, and the NVIDIA driver is rechecked each time. The author states that when Torch is upgraded to 2.14 and CUDA to 13.4, the installer will be updated accordingly.

Applicable Boundaries and Caveats

This toolset is aimed at users who want to run AI applications locally but do not want to handle CUDA and Python dependencies manually. A PC without an NVIDIA GPU can still complete all installation and compilation tests, but CUDA applications require an NVIDIA GPU to run. Windows 11 is recommended (Windows 10 21H2+ is compatible), and Ubuntu 24.04 is recommended (22.04 and 26.04 are compatible).

It is worth noting that the installer creates a 100GB page file/swap file, so systems with tight disk space should evaluate this in advance. In addition, although the tests cover a wide range, actually running AI applications also depends on the dependency installation of the specific application; the installer solves the system-level foundation, not application-level dependencies.

Obtaining the installer requires going through the author's Patreon page, as it is paid or membership content. For users who frequently tinker with local AI environments or need to configure multiple machines uniformly, the time savings from such a tool has real value; for users who occasionally run a single application, the investment needs to be weighed.

One-line Summary

SECourses releases one-click AI environment installers for Windows/Ubuntu that automate the full toolchain including CUDA and Python and pass nearly a hundred real compilation tests.

Paragraph Summary

This article introduces the one-click AI environment installers released by SECourses, designed to solve the pain points of complex and conflict-prone configuration of CUDA, cuDNN, multiple Python versions, and the C++ toolchain when running AI applications locally. On Windows, a double-clicked bat script performs a 20-step automated installation; on Ubuntu, a single sudo command completes it, covering NVIDIA drivers, CUDA 12.9/13.2, cuDNN 9.20, Python 3.10-3.14, Visual Studio Build Tools, FFmpeg, and other components. The installer does not just check version numbers; it runs 78 (Windows) and 66 (Ubuntu) compilation and execution tests in the actual environment of a new terminal, including GPU computation, CUDA extension builds, and FFmpeg encoding. The PATH order is specially tuned, incremental updates, automatic backups, and one-click restore are supported, and it can be run repeatedly. The tool solves the system-level runtime foundation without replacing application-level dependencies, and it must be obtained through Patreon for a fee.