"CUDA driver version is insufficient for CUDA runtime version"
Your NVIDIA driver is older than the CUDA build of PyTorch (or another library) needs. Either update the driver, or install a build made for an older CUDA version.
What you see
RuntimeError: CUDA error: CUDA driver version is insufficient for CUDA runtime versionor, from PyTorch itself:
RuntimeError: The NVIDIA driver on your system is too old (found version 12040). Please update your GPU driver ...The "found version" number is not your driver version. It's the newest CUDA version your driver supports, written as 1000 × major + 10 × minor: 12040 means CUDA 12.4, 11080 means CUDA 11.8. nvidia-smi shows the same thing as "CUDA Version" in its header, next to the actual "Driver Version".
A related message, Found no NVIDIA driver on your system, means no driver is loaded at all. That's an installation problem, not a version one; see torch.cuda.is_available() returns False.
Which driver each PyTorch 2.14.0 build needs
| CUDA build | Full requirement (Linux) | Full requirement (Windows) | Runs from (minor version compatibility) |
|---|---|---|---|
| CUDA 12.6 | 560.28.03 | 560.76 | 525+ |
| CUDA 13.0 plain pip install torch on Linux | 580.65.06 | not published | 580+ |
| CUDA 13.2 | 595.45.04 | not published | 580+ |
Minor version compatibility (NVIDIA's term) means a CUDA 12.x build runs on any driver from 525, and a 13.x build on any driver from 580, even below that exact version's own requirement. The exception is code compiled to PTX at runtime, which in PyTorch meanstorch.compile and Triton kernels: those need the full requirement. So if basic PyTorch works but torch.compile fails, your driver is between the two numbers.
The fix
- Update the driver if you can. That also unlocks the plain
pip install torch, which needs the default build's driver. - Or install an older CUDA build your driver supports, e.g.
pip install torch==2.14.0 --index-url https://download.pytorch.org/whl/cu126. PyTorch bundles its own CUDA runtime, so you don't need a matching system CUDA Toolkit; only the driver matters. - Under WSL2, update the driver on Windows, not inside Linux.
The install checker takes your driver version and GPU and picks the newest build that works, telling you which driver update would unlock a newer one.
Sources: minimum drivers from NVIDIA's CUDA release notes (see CUDA versions); minor version compatibility from NVIDIA's CUDA Compatibility guide (13.x ≥ 580, 12.x ≥ 525, 11.x ≥ 450); error text from PyTorch's c10/cuda/CUDAFunctions.cpp and the CUDA runtime.