PyTorch CUDA error: driver version is insufficient for CUDA runtime

PyTorch was built against a newer CUDA toolkit than your installed NVIDIA driver supports. Each driver has a maximum CUDA runtime version it can run; nvidia-smi's "CUDA Version" header shows that maximum, NOT the installed toolkit version.

Common scenario: pip install torch pulled the latest cu126 wheel; your driver is 535 (max CUDA 12.2). The wheel's runtime won't talk to the older driver.

1. Read the actual driver and runtime versions:

nvidia-smi  # "CUDA Version" = max runtime supported
python -c "import torch; print(torch.version.cuda)"  # actual runtime PyTorch wants

2. Easier path: downgrade PyTorch to match your driver:

pip uninstall torch torchvision -y

# Driver supports up to CUDA 12.1 → install cu121 wheel
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

PyTorch publishes wheels for CUDA 11.8, 12.1, 12.4, 12.6, 12.8.

3. Harder path: update the driver. Linux:

# Ubuntu — pick latest stable
sudo apt install nvidia-driver-560
sudo reboot

Windows: download from nvidia.com/Download. WSL2 users update the Windows host driver, not anything inside WSL.

4. Confirm after the fix:

python -c "import torch; print(torch.cuda.is_available())"  # True