PyTorch Installation Guide – GPU Setup & Troubleshooting

1. What is PyTorch? What you’ll learn in this article

What is PyTorch?

What you’ll learn in this

1. How to verify the required environment and tools before installing PyTorch.
2. Criteria for deciding whether to use pip or Anaconda.</>
3. Step-by-step installation procedures and configuration.
4. Troubleshooting and error resolution approaches.
5. Next steps and recommended resources after installation.

Intended Audience

• Beginners interested in deep learning but confused about the PyTorch setup process.
• Intermediate users who understand basic Python operations but lack confidence in library management or environment setup.
• Engineers or data scientists who want to build high-performance computing environments using GPUs.

2. Preparation before installation: required environment and tools

Checking the operating environment

1. OS support status

• Windows 10 or later
• macOS 10.15 (Catalina) or later
• Linux (Ubuntu, CentOS, and other major distributions)

2. Python version

python --version

python3 --version

3. Checking pip version

pip --version

python -m pip install --upgrade pip

Checking GPU environment (optional)

1. Verifying CUDA‑compatible GPU

nvidia-smi

2. Checking CUDA and cuDNN versions

nvcc --version

3. Installing CUDA

Note: When GPU is unavailable

Setting up a virtual environment (recommended)

1. When using venv

python -m venv pytorch-env
source pytorch-env/bin/activate  # Linux/MacOS
pytorch-envScriptsactivate     # Windows

2. When using Anaconda

conda create -n pytorch-env python=3.10
conda activate pytorch-env

Next steps

3. Comparison of Installation Methods and Selection Criteria

pip vs Anaconda: Which Should You Choose?

1. Installing with pip

• Easy and fast installation.
• Install only the necessary libraries with minimal overhead.
• Easy to combine with virtual environments (venv).

• Dependency management can become complex.
• GPU driver and CUDA compatibility must be managed manually.

• For lightweight scripts or projects.
• When you want to add PyTorch to an existing Python environment.

2. Installing with Anaconda

• Intuitive and easy virtual environment management.
• Automatic dependency resolution reduces installation issues.</>
• Works well with Jupyter Notebook and data analysis tools.

• Large installation size and longer setup time.
• Commands may feel complex for beginners.

• Ideal for machine learning and data analysis projects.
• When you are concerned about dependency or package management.

3. Summary of Selection Criteria

4. Which Should You Choose?

• When you want a quick try: pip is recommended. It can be added quickly to an existing Python environment, making it ideal for those who want to start with a minimal setup.
• When building a full‑scale project: Anaconda is recommended. It is optimal for cases that require robust dependency management and large‑scale environment setup.

Next Steps

4. PyTorch Installation Guide (Practical Edition)

1. Install with pip (Fastest Method)

1-1. CPU Version Installation

pip install torch torchvision torchaudio

1-2. GPU Version Installation (CUDA Support)

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

1-3. Verify Installation

import torch
print(torch.__version__)  # display version
print(torch.cuda.is_available())  # check GPU support

2.0.1+cu118
True

2. Install with Anaconda (Recommended Method)

2-1. Create and Activate a Virtual Environment

conda create -n pytorch-env python=3.10
conda activate pytorch-env

2-2. CPU Version Installation

conda install pytorch torchvision torchaudio cpuonly -c pytorch

2-3. GPU Version Installation (CUDA Support)

conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

2-4. Verify Installation

import torch
print(torch.__version__)  # display version
print(torch.cuda.is_available())  # check GPU support

2.0.1+cu118
True

3. Troubleshooting

3-1. Dependency Errors with pip

ERROR: Could not find a version that satisfies the requirement torch

pip install --upgrade pip

3-2. Package Resolution Errors with Anaconda

PackagesNotFoundError: The following packages are not available

4. Best Practices for Environment Setup

• Use Virtual Environments: Separate environments per project to avoid dependency conflicts.
• Package Management Tips: Do not mix pip and conda; maintain consistency per environment.
• Record Dependencies: Export and back up dependencies after installation.

pip freeze > requirements.txt

conda list --export > environment.yml

Next Steps

5. Setting up CUDA and cuDNN (GPU usage)

1. What is CUDA?

2. What is cuDNN?

1. CUDA installation steps

1-1. Verify GPU

nvidia-smi

1-2. Verify CUDA compatibility

1-3. Download and install CUDA

1. From NVIDIA’s CUDA download page, select the OS and version you need.
2. Follow the instructions to download the installer and run the installation.

1-4. Verify installation

nvcc --version

Cuda compilation tools, release 11.8, V11.8.89

2. cuDNN installation steps

2-1. Download cuDNN

1. Download from NVIDIA Developer site’s cuDNN page.
2. An NVIDIA account is required; log in or register.

2-2. Installation steps

1. Extract the downloaded ZIP file.
2. Copy the extracted files into the CUDA installation directory.

sudo cp cuda/include/* /usr/local/cuda/include/
sudo cp cuda/lib64/* /usr/local/cuda/lib64/

2-3. Set environment variables

export PATH=/usr/local/cuda/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH

2-4. Verify configuration

cat /usr/local/cuda/include/cudnn_version.h | grep CUDNN_MAJOR -A 2

#define CUDNN_MAJOR 8
#define CUDNN_MINOR 0
#define CUDNN_PATCHLEVEL 5

3. Test PyTorch and CUDA compatibility

3-1. Verify with Python code

import torch
print(torch.cuda.is_available())  # If True, CUDA is recognized
print(torch.cuda.get_device_name(0))  # Display GPU model name

True
NVIDIA GeForce RTX 3060

3-2. Troubleshooting errors

torch.cuda.is_available() -> False

• Check the version compatibility between CUDA and PyTorch and install the appropriate versions.
• Update the NVIDIA driver to the latest version.

RuntimeError: CUDA error: invalid device function

• If the CUDA version is outdated, update to the latest version.
• Also verify the cuDNN version and ensure compatibility.

Next steps

6. Verifying the Installation

1. Verifying the Installation with Python Code

1-1. Basic Functionality Check

import torch
print("PyTorch Version:", torch.__version__)  # version check

PyTorch Version: 2.0.1

1-2. Checking GPU Support

print("CUDA Available:", torch.cuda.is_available())  # GPU support availability

CUDA Available: True

1-3. Checking Detailed GPU Information

if torch.cuda.is_available():
    print("Device Name:", torch.cuda.get_device_name(0))  # display GPU model name
    print("CUDA Version:", torch.version.cuda)  # display CUDA version

Device Name: NVIDIA GeForce RTX 3060
CUDA Version: 11.8

2. Running Simple Code for Verification

Tensor Generation and Computation

import torch

# generate a random 3x3 matrix
x = torch.rand(3, 3)
print("Random Tensor:")
print(x)

# matrix addition
y = torch.rand(3, 3)
z = x + y
print("Addition Result:")
print(z)

Random Tensor:
tensor([[0.2567, 0.8954, 0.6543],
        [0.1234, 0.4567, 0.7890],
        [0.9876, 0.5432, 0.1123]])

Addition Result:
tensor([[0.6789, 1.3456, 1.1234],
        [0.3456, 0.7890, 1.5678],
        [1.4321, 0.9876, 0.5432]])

3. Common Errors and Solutions

3-1. When PyTorch Cannot Be Imported

ModuleNotFoundError: No module named 'torch'

1. Reactivate the virtual environment.

conda activate pytorch-env

2. Reinstall it.

pip install torch torchvision torchaudio

3-2. When the GPU Is Not Recognized

CUDA Available: False

• CUDA is not installed.
• The CUDA version is incompatible with PyTorch.
• The GPU driver is outdated.

1. Check the NVIDIA driver version and update to the latest.

nvidia-smi

2. Check the CUDA version.

nvcc --version

3. Reinstall drivers or CUDA as needed.

4. Troubleshooting Tips

1. Refer to the official documentation: If a problem occurs, consult the PyTorch official support page.
2. Search the error message: Googling the exact error message often leads to solutions on forums or GitHub Issues.
3. Align versions: Compatibility among CUDA, cuDNN, and PyTorch versions is crucial. Refer to the official compatibility matrix to verify the latest versions.

Next Steps

7. Common Issues and Solutions

1. Installation Errors and Solutions

1-1. Dependency Mismatch Error

ERROR: Could not find a version that satisfies the requirement torch

• The pip or Python version is outdated, causing incompatibility with packages required by PyTorch.
• The specified CUDA version is not compatible with the selected PyTorch version.

1. Update pip and Python to the latest versions.

python -m pip install --upgrade pip

2. Recheck the environment’s dependencies.

pip check

3. Obtain the appropriate installation command from the official site and try again.

1-2. Virtual Environment Errors

ModuleNotFoundError: No module named 'torch'

1. Reactivate the virtual environment.

conda activate pytorch-env

2. Recreate the virtual environment.

conda create -n pytorch-env python=3.10
conda activate pytorch-env
conda install pytorch torchvision torchaudio -c pytorch

2. GPU Configuration Errors and Solutions

2-1. GPU Not Detected

torch.cuda.is_available() -> False

• CUDA is not installed correctly.
• The GPU driver version is outdated.
• The PyTorch and CUDA versions are incompatible.

1. Check the NVIDIA driver version and update to the latest.

nvidia-smi

2. Verify the CUDA version.

nvcc --version

3. Reinstall the driver and CUDA if necessary.

2-2. CUDA Error

RuntimeError: CUDA error: invalid device function

1. Check the official compatibility matrix for CUDA and PyTorch versions.
2. Install using a compatible version.

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

3. Runtime Errors and Solutions

3-1. Out-of-Memory Error

RuntimeError: CUDA out of memory.

1. Reduce the batch size.
2. Delete unnecessary variables to free memory.

del variable_name
torch.cuda.empty_cache()

3. Monitor memory usage and optimize the code accordingly.

3-2. Slow Performance

1. Check whether the GPU is recognized.

print(torch.cuda.is_available())

2. Explicitly specify the device in the code.

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)

4. Other Issues and Solutions

4-1. When a PyTorch Update Is Needed

pip install --upgrade torch torchvision torchaudio

4-2. Using Support Forums

• Official Documentation: See the support page on the PyTorch official site.
• Community Forums: Ask questions on Stack Overflow or PyTorch Discuss.

Next Steps

8. Summary and Next Steps

1. Review of This Article’s Key Points

1. What is PyTorch?

• A framework optimized for machine learning and deep learning, supporting fast GPU computation.

2. Pre‑Installation Preparation

• You need to verify the OS and Python versions and prepare for CUDA and cuDNN installation.
• Using a virtual environment is recommended to simplify dependency management.

3. Installation Procedures

• pip: Lightweight and quick to install.
• Anaconda: Easy environment management, suitable for large‑scale projects.

4. Configuring CUDA and cuDNN

• If you plan to use a GPU, you must verify compatibility and install CUDA and cuDNN.

5. Verifying Installation and Troubleshooting

• After installation, check the version and GPU support, and we also provide solutions for any errors that arise.

2. Next Steps Using PyTorch

2-1. Building Your First Neural Network

import torch
import torch.nn as nn
import torch.optim as optim

# Prepare data
x = torch.rand(100, 1)
y = 2 * x + 1 + 0.1 * torch.randn(100, 1)

# Define model
model = nn.Linear(1, 1)
loss_fn = nn.MSELoss()
optimizer = optim.SGD(model.parameters(), lr=0.01)

# Training loop
for epoch in range(1000):
    y_pred = model(x)
    loss = loss_fn(y_pred, y)
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

# Verify results
print(model.weight.item(), model.bias.item())

2-2. Applied PyTorch Projects

1. Creating an Image Classification Model

• Practice an image classification task using a CNN (Convolutional Neural Network).

1. Building a Natural Language Processing Model

• Try text classification or translation models using RNNs or Transformers.

1. Leveraging Pre‑trained Models

• Build advanced applications using pre‑trained models such as ResNet or BERT.

3. Learning Resources

3-1. Official Documentation and Tutorials

• PyTorch Official Site
• PyTorch Tutorials

3-2. Online Courses and Video Learning

• PyTorch courses are available on online learning platforms such as Coursera and Udemy.
• YouTube also offers free practical videos, which are recommended.

3-3. Hands‑On Projects and Sample Code

• • Referencing projects and sample code published on GitHub can help you learn practical usage.