Skip to main content

nvidia-smi shows CUDA 12.2, but nvcc --version doesn't work, and PyTorch doesn't support 12.2

Problem​

nvidia-smi shows CUDA 12.2, but PyTorch or nvcc shows a different version or doesn't run.

Cause​

The CUDA version shown by nvidia-smi is the maximum CUDA version supported by the GPU driver. That is different from the actual CUDA Toolkit (nvcc) version and the CUDA version your framework uses.

Solution​

  1. Check the actual CUDA Toolkit version with:
nvcc --version
  1. If the command isn't recognized, set your PATH:
export PATH=/usr/local/cuda/bin:$PATH
nvcc --version
  1. Install a PyTorch / TensorFlow build that matches the CUDA Toolkit version you confirmed.

Additional Notes​

  • PyTorch checks compatibility based on the CUDA Toolkit, not nvidia-smi.

  • Example: nvidia-smi → 12.2, nvcc → 11.8 → use PyTorch cu118.