CUDA version for fastbook

Part 1 (2020) Non-beginner
1

Hi fastaiers,

I am setting up a Ubuntu 20.04 box, with a GTX 1080 Ti, to work through the fastbook - excited :slight_smile: !

Is there a specific CUDA version that is required or recommended?

In fact, are there recommended versions for:

  • Python
  • PyTorch
  • CUDA
  • CUDNN

Cheers

2

Hello there,

3.9 is the latest stable version of Python, but I use 3.10 and have not thus far encountered any errors with PyTorch or fastai. However, note that some libraries like TensorFlow only support up to 3.9, so that’s what you might be better off with.

The nice thing about PyTorch is that it takes care of CUDA dependencies for you (regardless of whether your package manager is pip or Anaconda), therefore no need to worry about them. You can check out the PyTorch docs for more info on that.

Oh, and fastai recently announced fastchan, a “conda mini-distribution with a focus on the PyTorch ecosystem” that makes the “installation and updates of libraries such as PyTorch and RAPIDS […] faster, easier, and more reliable”. Considering your situation, it may be worthwhile to give it a look.

Cheers!

1 Like
3

Thanks, that is helpful. I’ve been using Python 3.7 so far because the Miniconda default 3.8 gave me some trouble with CUDA (it somehow did not detect my GPU). I didn’t think to try 3.9 or 3.10, but I’ll give those a shot!

1 Like
5

So, I’ve been trying to use Python 3.10 and I have had nothing but pain for my trouble.

I finally tried using mamba to manage the packages. I think the problem boils down to this:

package fastai-2.5.2-py_0 requires pytorch >=1.7.0,<1.10, but none of the providers can be installed

Based on another thread and this open development issue. It looks like it’s being actively worked on by Pytorch developers, but they note it may take a bit longer to get conda-friendly packages out there.

Not sure what dark magic @BobMcDear employed to get Python 3.10 to work, but then, I’m still pretty newbish when it comes to the wonderful world of Python package management. In any event, it looks like it might be solved in time. For now, I’ll learn to live without the niceties of Python 3.10. :crying_cat_face: