Captured at (local ISO): 2026-05-18 05:17:02

Preface

In deep learning, PyTorch is a top choice thanks to its flexible design and friendly API. Efficient PyTorch use depends on GPU drivers, CUDA (NVIDIA’s compute platform), and cuDNN (CUDA acceleration libraries) working together—missing any piece can mean “GPU not detected” or code falling back to CPU.
Whether on a local Windows/Linux machine or cloud servers (Alibaba Cloud, Tencent Cloud, etc.), environment setup is the first hurdle: incompatible driver versions fail installs; mismatched CUDA and PyTorch trigger “CUDA not found”; skipping cuDNN slows training a lot. This tutorial walks through the full install from zero, with verification at each step, helping you avoid most common pitfalls.

1. GPU Driver Install (local / cloud same logic)

Step 1: Check GPU model

Windows
Right-click desktop → open “NVIDIA Control Panel” → left “System Information” → “Display”, note the GPU model (e.g. RTX 4090, GTX 1660).

Screenshot limits made me photograph the screen with a phone
After opening:

Click “System Information” at the bottom left

The red box shows the exact GPU info

Linux
After SSH to the server, run:

lspci | grep -i nvidia  # e.g. "NVIDIA Corporation GA102 [GeForce RTX 3090]"

If no GPU appears on cloud, confirm the instance type includes GPU (e.g. Alibaba “GN6”, Tencent “GT4” series).

Step 2: Download and install driver

1. Download from NVIDIA
Open the NVIDIA driver lookup site (https://www.nvidia.cn/drivers/lookup/)
Pick options for your GPU—for example:
“Product Type”: “Data Center / Tesla” (cloud server GPUs);
“Product Series / Model”: Tesla A10;
“Operating System”: “Windows Server 2025” (chosen when provisioning the server);
Click “Search” → “Download” for the .exe.


On the results page, review and download

2. Install driver

1. Choose “Express install” (recommended for beginners);
2. Prefer default install path;
3. If prompted about an old driver, choose “overwrite”;
4. Reboot after install.

Step 3: Verify install

Press Win+R, type cmd, Enter. In the terminal run nvidia-smi. Output like below means the driver is installed:

2. CUDA Install

CUDA is NVIDIA’s toolkit for GPU computing; PyTorch uses it to run on the GPU. Rule: CUDA version ≤ highest CUDA version your driver supports (e.g. driver shows “CUDA Version: 12.4” → you can install CUDA 12.4, 12.3, etc.).

Step 1: Check supported CUDA version

“CUDA Version” is the maximum supported—install CUDA below that (e.g. below 12.8 here).

Step 2: Download matching CUDA

Download from the archive: https://developer.nvidia.com/cuda-toolkit-archive

I used CUDA 12.1.0 to match my local machine and because very new CUDA versions are harder to pair with PyTorch builds.

Step 3: Install CUDA

Run the downloaded executable

Click through the installer with defaults.

Step 4: Configure environment variables

Before testing, set CUDA environment variables—the installer sets some, but not all.

1. Open the CUDA install path (often under C:\Program Files, look for CUDA folders)
   
   Open the folder with “toolkit” to see CUDA
   
   Note these paths (relative below—find the full absolute paths on your machine):
   **1. \bin
2. \lib\x64
3. \libnvvp
4. The toolkit root you have open**

Next open “Edit the system environment variables” (search in Windows).

Click Environment Variables

Edit PATH under System variables (bottom), not user variables

Then:

Click New to add entries. The installer may have added some—add any of the four paths that are missing.

Step 5: Verify install

Open a command prompt and run:

ncvv -V

Success screen looks like this

3. cuDNN Install

cuDNN is NVIDIA’s deep-learning library; it speeds PyTorch/TensorFlow on CUDA and must match your CUDA version exactly.
Download from https://developer.nvidia.com/rdp/cudnn-archive

Pick a build that matches your CUDA major version (12.x covers 12.1, 12.2, etc.).

Extract the zip—you get three folders

Open your CUDA install directory and copy files from the cuDNN folders into the matching CUDA folders:
bin → bin
include → include
lib/x64 → lib/x64

After copying, cuDNN is installed.

4. PyTorch Install

PyTorch installs like any Python package; mirrors help. For pip mirrors see:
Python: pip too slow? Two steps to configure a domestic mirror
Nanjing University’s mirror is recommended—full PyTorch wheels and frequent updates: https://mirrors.nju.edu.cn
If a version is “not found”, check pip syntax; you can install from a direct wheel URL:

pip install <paste the wheel URL from the mirror>

Example: finding 2.5.0+cu121 on the mirror:
Add pytorch to the NJU mirror base URL → https://mirrors.nju.edu.cn/pytorch

Open whl → https://mirrors.nju.edu.cn/pytorch//whl

cpu = CPU-only PyTorch (no GPU)—slower
cuXXX = CUDA builds; we used CUDA 12.1, so open cu121

Find the torch folder

Pick your Python version and platform, copy the link, pip install it.

Verify:

pip show torch

Success looks like:

5. torchvision Install

torchvision handles image data (datasets, transforms) and must match your PyTorch version.
For PyTorch 2.5.1+cu121, use torchvision 0.20.1 per official compatibility. Find it on the same mirror or:

pip install torchvision==0.20.1+cu121 --index-url https://download.pytorch.org/whl/cu121

Verify:

pip show torchvision

Success:

Summary

The core rule is version matching: GPU driver → CUDA → cuDNN → PyTorch → torchvision, each step aligned. Local and cloud logic are the same; only Linux/Windows commands and paths differ.
If something fails, check: ① driver supports your CUDA; ② nvcc -V finds CUDA; ③ PyTorch install targets the right CUDA tag. Follow this guide and you should get a working stack in one pass.