GPU & AI in WSL

Unlock GPU passthrough, CUDA, and machine-learning frameworks inside WSL 2 for local AI development.

GPU Passthrough in WSL 2

WSL 2 supports GPU paravirtualization through the vGPU driver model. Your Windows GPU driver automatically exposes the GPU inside every WSL distro — no extra guest driver required.

NVIDIA: Install the latest Game Ready or Studio driver on Windows — CUDA support is included automatically.

AMD: Use the Adrenalin driver with DirectML support for inference workloads.

Intel: Arc and integrated GPUs work via the Intel GPU driver with DirectML and oneAPI support.

Ubuntu (WSL)
# Verify GPU is visible inside WSL
user@wsl:~$ nvidia-smi
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 560.x Driver Version: 560.x CUDA Version: 12.6 |
| GPU Name ... |
| 0 NVIDIA GeForce RTX 4090 ... |
+-----------------------------------------------------------------------------------------+

CUDA Toolkit Installation

With the Windows driver providing GPU access, you only need the CUDA toolkit inside WSL — not a separate Linux GPU driver.

Important: Do not install a separate NVIDIA Linux driver inside WSL. The Windows host driver handles all GPU communication. Installing a Linux driver will break GPU passthrough.

Ubuntu (WSL)
# Add NVIDIA CUDA repository
user@wsl:~$ wget https://developer.download.nvidia.com/compute/cuda/repos/wsl-ubuntu/x86_64/cuda-wsl-ubuntu.pin
user@wsl:~$ sudo mv cuda-wsl-ubuntu.pin /etc/apt/preferences.d/cuda-repository-pin-600
# Install CUDA toolkit (without the driver)
user@wsl:~$ sudo apt update && sudo apt install -y cuda-toolkit-12-6
# Add CUDA to your PATH
user@wsl:~$ echo 'export PATH=/usr/local/cuda/bin:${PATH}' >> ~/.bashrc
user@wsl:~$ echo 'export LD_LIBRARY_PATH=/usr/local/cuda/lib64:${LD_LIBRARY_PATH}' >> ~/.bashrc
Ubuntu (WSL)
# Verify CUDA installation
user@wsl:~$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Cuda compilation tools, release 12.6, V12.6.x

PyTorch & TensorFlow Setup

Both frameworks detect CUDA inside WSL automatically. Use a virtual environment to keep your system Python clean.

PyTorch with CUDA

Ubuntu (WSL)
# Create a virtual environment
user@wsl:~$ python3 -m venv ~/ai-env && source ~/ai-env/bin/activate
# Install PyTorch with CUDA support
(ai-env) user@wsl:~$ pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
# Verify GPU is available
(ai-env) user@wsl:~$ python3 -c "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0))"
True NVIDIA GeForce RTX 4090

TensorFlow with CUDA

Ubuntu (WSL)
# Install TensorFlow (GPU support included by default)
(ai-env) user@wsl:~$ pip install tensorflow
# Verify GPU detection
(ai-env) user@wsl:~$ python3 -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

Jupyter Notebooks in WSL

Run Jupyter Lab inside WSL and access it from your Windows browser. VS Code's Jupyter extension also connects to WSL kernels seamlessly.

Ubuntu (WSL)
# Install JupyterLab
(ai-env) user@wsl:~$ pip install jupyterlab ipykernel
# Register the virtual environment as a Jupyter kernel
(ai-env) user@wsl:~$ python3 -m ipykernel install --user --name=ai-env --display-name "AI Environment"
# Launch JupyterLab (opens in Windows browser automatically)
(ai-env) user@wsl:~$ jupyter lab --no-browser --ip=0.0.0.0 --port=8888
http://127.0.0.1:8888/lab?token=abc123...

VS Code tip: Open VS Code with the Remote - WSL extension, then open any .ipynb file. VS Code will detect your Jupyter kernels inside WSL automatically.

Performance Tips

Store datasets on the Linux filesystem

Keep training data under /home/user/data instead of /mnt/c/. Cross-filesystem I/O is significantly slower and can bottleneck data-loading pipelines.

Allocate enough memory in .wslconfig

Large models need RAM. Set memory=16GB (or more) in %USERPROFILE%\.wslconfig to avoid out-of-memory crashes during training.

Use Docker with GPU support

Docker Desktop with the WSL 2 backend passes through GPUs. Use docker run --gpus all to run containers with CUDA access — great for reproducible ML pipelines.

Next Steps

Development Setup

Set up full dev environments with Docker, VS Code, and language runtimes.

WSL Configuration

Fine-tune .wslconfig for memory, CPU, and networking to optimize GPU workloads.

Networking Deep Dive

Configure port forwarding to expose Jupyter and model endpoints to your network.

Troubleshooting

Fix common GPU detection issues and CUDA version mismatches.