Wsl2 Gpu Cuda

WSL2 GPU / CUDA

Enabling GPU acceleration (CUDA, Vulkan, DirectML) in WSL2.

Prerequisites

# 1. Windows 11 or Windows 10 21H2+
# 2. WSL2 with GPU paravirtualization
wsl --set-default-version 2

# 3. Install NVIDIA drivers on Windows (covers WSL2 too)
# Download from: https://www.nvidia.com/Download/index.aspx
# DO NOT install Linux drivers inside WSL2 -- Windows driver covers both

CUDA in WSL2

# Verify GPU is visible from WSL2
nvidia-smi

# Should show: Driver Version matches Windows driver
# CUDA Version: 12.x

# Run CUDA containers
docker run --gpus all nvidia/cuda:12.2-runtime-ubuntu22.04 nvidia-smi

Docker with GPU

docker run --gpus all -it nvidia/cuda:12.2-runtime-ubuntu22.04 bash
# Inside container: nvidia-smi should work

PyTorch/TensorFlow with GPU

# CUDA toolkit inside WSL2 (for native, not Docker)
wget https://developer.download.nvidia.com/compute/cuda/repos/wsl-ubuntu/x86_64/cuda-keyring_1.1-1_all.deb
sudo dpkg -i cuda-keyring_1.1-1_all.deb
sudo apt-get update
sudo apt-get install -y cuda-toolkit-12-2

# Verify
nvcc --version

Vulkan in WSL2

# Install Vulkan SDK and tools
sudo apt-get update
sudo apt-get install -y vulkan-tools mesa-vulkan-drivers

# Verify
vulkaninfo
vkcube

DirectML (for AMD/Intel GPUs)

# DirectML in WSL2 doesn't need special drivers
# Windows GPU drivers are forwarded

# For PyTorch with DirectML
pip install torch-directml

Performance Notes

# GPU in WSL2 is near-native performance for compute (CUDA)
# Graphics/display is paravirtualized -- slightly lower perf than native

# Monitor GPU usage
nvidia-smi -l 1
watch -n 1 nvidia-smi

Troubleshooting

# nvidia-smi: "command not found"
# Install NVIDIA CUDA tools on WSL2:
sudo apt-get install -y nvidia-utils-545

# nvidia-smi: "No devices were found"
# Check Windows drivers are installed
# Check `wsl --shutdown` then restart

# Docker --gpus all: "could not select device driver"
# Install nvidia-container-toolkit inside WSL2:
sudo apt-get install -y nvidia-container-toolkit
sudo systemctl restart docker

Pitfalls

• Windows GPU drivers cover WSL2 -- do NOT install Linux GPU drivers inside WSL2
• nvidia-smi shows Windows driver version, not a WSL2 version
• --gpus all requires nvidia-container-toolkit installed in the Docker host (WSL2)
• Not all CUDA versions are supported -- check NVIDIA's WSL2 compatibility matrix
• GPU compute works; GPU display/output may have limitations