HashNotes

Upgrade CUDA version from 11.6 to 11.8

Mohamad Mahmood's photo
Mohamad Mahmood
·

2 min read

Although the NVIDIA app is the best choice for upgrading the graphic driver, there are some cases where you may need to manually install drivers from the NVIDIA website:

  1. If you need a specific older driver (e.g., for legacy software compatibility).

  2. If NVIDIA app fails to detect or update the driver correctly.

  3. If you are using a workstation GPU (e.g., NVIDIA A100, Tesla, Quadro).

If your current CUDA version is 11.6, follow these steps to upgrade to CUDA 11.8 for better compatibility with PyTorch 2.x, Xformers, and Transformers.

Step 1: Check Current CUDA Version

Before upgrading, confirm your installed CUDA version:

nvidia-smi

If it shows CUDA 11.6, proceed to upgrade.

Step 2: Download and Install CUDA 11.8

  • Go to the NVIDIA CUDA Toolkit 11.8 download page: CUDA 11.8 Downloads

  • Select the following options:

    • Operating System: Windows

    • Architecture: x86_64

    • Version: 11.8

    • Installer Type: EXE (Local)

  • Download the CUDA 11.8 Installer (~3GB).

  • Run the Installer and select:

    • Custom Install

    • Uncheck "NVIDIA Graphics Driver" (To keep your existing drivers)

    • Install only CUDA and cuDNN

The error "No supported version of Visual Studio was not found" means that some CUDA components (e.g., NVCC compiler, cuDNN, CUDA libraries) require a supported version of Visual Studio.

However, CUDA will still work for PyTorch and TensorFlow, even if some components fail.
If you plan to compile CUDA code manually (e.g., custom kernels, C++ extensions), you must install Visual Studio.

cuda