WSL2: torch.cuda.is_available() returns False

Environment: Windows 11 + WSL2 (Ubuntu) running PyTorch / vLLM / llama.cpp.

Severity: high — GPU is invisible to the WSL distro.

• Windows host driver too old (must be ≥ 470.76 for CUDA-on-WSL)
• Recent Windows kernel update broke /dev/dxg passthrough (a known recurring issue)
• WSL kernel too old (wsl --update not run since Windows install)
• User installed Linux NVIDIA driver inside WSL — this conflicts with the host passthrough
• Distro is WSL1 not WSL2 (wsl -l -v to verify)
• Custom kernel without DXGKRNL support

1. Update WSL kernel from PowerShell (admin):

wsl --update
wsl --shutdown

Restart WSL afterwards.

2. Update the Windows NVIDIA driver to the latest Game Ready or Studio driver (≥ 535+):

# Verify from Windows side first
nvidia-smi

If that fails on Windows, fix host before touching WSL.

3. Verify /dev/dxg exists inside WSL:

ls -la /dev/dxg
# Should exist; if missing, kernel doesn't support GPU passthrough

4. If you installed Linux NVIDIA drivers by mistake, remove them:

sudo apt purge 'nvidia-*' libnvidia-*
sudo apt autoremove
# Reboot WSL: wsl --shutdown (from PowerShell)

5. Reinstall CUDA toolkit (NOT the driver) for compilation only:

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 update && sudo apt install -y cuda-toolkit-12-4

6. Confirm:

nvidia-smi  # Should show the GPU
python -c "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0))"

Caveat: if a recent Windows update broke GPU-on-WSL (this has happened multiple times in the 11-26100 series), the workaround is sometimes to roll back the Windows update or install an Insider Preview build with the fix. Watch the WSL GitHub issue tracker for current advisories.