CUDA driver version mismatch

What’s actually happening

Each CUDA toolkit version requires a minimum NVIDIA driver version. CUDA_ERROR_INSUFFICIENT_DRIVER means the toolkit your inference framework was built against expects driver capabilities your currently installed driver doesn’t provide — this is a driver-vs-toolkit version mismatch, not a GPU compatibility problem.

Check both versions

nvidia-smi

The top-right of nvidia-smi’s output shows your installed driver version and the maximum CUDA version that driver supports — this is different from the CUDA toolkit version your Python environment or inference framework has installed.

nvcc --version

Compare the CUDA version nvcc reports against the maximum version nvidia-smi said your driver supports. If nvcc’s version is higher, that’s the mismatch.

Fix

Update your NVIDIA driver to a version that supports your installed CUDA toolkit — check NVIDIA’s CUDA Toolkit release notes for the minimum required driver version for your toolkit release, since this mapping changes with each CUDA release and isn’t safe to assume.

On Linux, avoid mixing driver installation methods (distro package manager vs. NVIDIA’s .run installer vs. Docker’s NVIDIA Container Toolkit) — partial installs from different sources are a common cause of drivers reporting an unexpected version.

# Ubuntu/Debian: check what's actually installed
dpkg -l | grep nvidia-driver

If you’re running inference inside Docker, confirm the host driver (not a driver inside the container) meets the requirement — GPU passthrough uses the host’s driver, and installing a driver inside the container has no effect.