Cuda and pytorch versions

I am sorry if my question is answered before.

I am setting my personal Desktop for the course along with the AWS as I want to try both.

Can anyone mention will the course require cuda 8 or cuda 9? also which pytorch version is required?

Thanks

For PyTorch, you have the choice between CUDA v7.5 or 8.0. Be sure to install the right version of cuDNN for your CUDA. The most recent version of PyTorch is 0.2.0_4.

2 Likes

If you go to http://pytorch.org/, you will see that they have binaries for only CUDA 7.5 or 8. So choose the right one based on your hardware and what Nvidia Driver you installed. Checkout this blog for any reference on CUDA installation - http://www.pradeepadiga.me/blog/2017/03/22/installing-cuda-toolkit-8-0-on-ubuntu-16-04/

Thanks for your replies.
What I meant is the specific versions that will be used for the course not the available versions.
Also note you could install pytorch with cuda 9.0 from binaries.

We’re using the latest pytorch

1 Like

Thank you.

I’m going the Docker way. Nvidia Docker v2 is simple to install and use. It only needs Nvidia GPU drivers to be installed CUDA is supplied by Nvidia-docker. I faced problems running different combinations of Cuda and CuDnn versions in the past. Hopefully this will also keep my work more reproducible.

3 Likes

I have just set up my laptop which has a GTX 1070. I wanted to have both pytorch and tensorflow. The specs on the tensorflow installation page for ubuntu works well for me (https://www.tensorflow.org/install/install_linux). Currently, the best way to go is cuda 8 and cudnn 6. cuda 9 and cudnn 7 are out there but pytorch is not yet generally compatible with cuda 9 and tensorflow does not work with cudnn 7 (don’t know about pytorch).

1 Like

yes you are right, I will go with cuda 8 , cudnn 6 and latest pytorch release.
I think pytorch comes with its cudnn, you just need cuda for pytorch. Correct me if I am wrong.

@visnaga Not entirely sure to be honest, you may well be right.

According to soumith:

PyTorch binaries don’t rely on a system CUDA being installed, they ship relevant bits themselves. You don’t need to install CUDA or cuDNN. All you need to use PyTorch on the GPU is a working NVIDIA driver. https://github.com/pytorch/pytorch/issues/4156#issuecomment-351512834

I just installed pytorch on my Ubuntu desktop, and it seems CUDA 9 is now supported.

There is now the NVIDIA docker with everything pre-installed!

You just need to NVIDIA drivers for your graphics card, no need to install CUDA or cuDNN anymore.

  1. Install Docker: https://docs.docker.com/install/
  2. Install nvidia-docker: https://github.com/NVIDIA/nvidia-docker
  3. Go to NVIDIA Cloud: https://ngc.nvidia.com/catalog/all
  4. Choose the PyTorch container
  5. Read the documentation to choose the right container!
  6. For example (at the time of writing): docker pull nvcr.io/nvidia/pytorch:19.12-py3
  7. Run! docker run --gpus all -it --rm --ipc=host -v /home/steeve/Data:/workspace/data nvcr.io/nvidia/pytorch:19.12-py3