Nvidia-drivers, cuda version and pytorch version compatibility

vision
1

I have 4 A100 graphics cards in the lab

  • GPU driver is 470.256.02
  • cuda version is 11.4
  • pytorch version is 1.13.
    But now I want to use functions such as torch.compile() which need pytorch verision >2.0.0. Since the GPU driver in the lab cannot be updated, the GPU driver is still 470.256.02.
  1. One way is to install cuda 11.8. Although the nvidia official website states that GPU drivers >450 are compatible with 11.x versions of cuda, some functions are lost. I am not sure whether there will be compatibility issues if I use GPU driver 470.256.02, cuda version 11.8 and pytorch version 2.1.
  2. The second method is to keep cuda version 11.4 and directly install pytorch 2.1 that supports cuda 11.8.
    I would like to ask if there is a better way to use the latest version of pytorch without updating the graphics card driver.
2

Your second approach should work, but I would highly recommend updating the drivers to any one supporting CUDA 12.x applications as the drivers were released at the beginning of 2023 already.