23.06 Install Page Update

[jarmak-nv]

This turned out smaller than expected, probably could have been part of #394 but I like smaller PRs anyway.

• Removed the mention of Python 3.8 support in pip installation guide
• Added CUDA 12.0 as a supported CUDA
• Added a note that the Docker/Conda CUDA 11 installations work fine on CUDA 12 machines, but that pip needs to match

Contributes to #386

[jarmak-nv]

@bdice Quick question on the bolded part below for this line since I think you're the person to ask (apologies if not!):

Follow this helpful developer guide{: target="_blank"} and then install the CUDA Toolkit without drivers{: target="_blank"} into the WSL2 instance. It's important to execute sudo apt-get -y install cuda-toolkit instead of sudo apt-get -y install cuda to avoid installing a GPU driver into WSL2. The Windows host system provides the driver to WSL2.

If they follow the steps from the link, is this accurate? We link to the guide to do a local install of a specific WSL Ubuntu 11.8 CTK that does not include the drivers.

The guide says to end with the command sudo apt-get -y install cuda rather than cuda-toolkit. If that's wrong, we should reach out to get that fixed too.

Relevant since I'll need to add the CUDA 12 link and just want to make sure it's correct. FWIW when I installed it on my WSL2 machine a couple months ago I followed the guide exactly without issue.

[bdice]

@jarmak-nv I want to be sure that users don't get tripped up by the second link. It says:

https://developer.nvidia.com/cuda-11-8-0-download-archive?target_os=Linux&target_arch=x86_64&Distribution=WSL-Ubuntu&target_version=2.0&target_type=deb_local

Note the last line, sudo apt-get -y install cuda is incorrect and should be cuda-toolkit. Otherwise WSL users will get a driver which is not desired.

[bdice]

If that's wrong, we should reach out to get that fixed too.

I'm not sure how to move that forward -- but I think the download page should never advise installing a driver into a WSL system. If you know where to ask, please inquire about that and feel free to tag me!

[jarmak-nv]

Don't the previous steps in the instructions make it install the version they downloaded though? the cuda-repo-wsl-ubuntu-11-8-local_11.8.0-1_amd64.deb (and any WSL-Ubuntu version) doesn't include a driver.

Per NV docs

We recommend developers to use a separate CUDA Toolkit for WSL 2 (Ubuntu) available from the CUDA Toolkit Downloads page to avoid this overwriting. This WSL-Ubuntu CUDA toolkit installer will not overwrite the NVIDIA driver that was already mapped into the WSL 2 environment.

I could easily be missing something though!

[bdice]

@jarmak-nv Thanks for pointing that out. I don't think I was aware of that difference until now. Maybe I'm concerned about something that has already been solved? I know that users (and my former self) have run into situations where drivers were installed in WSL but maybe it's a case where they (and I) haven't read the directions carefully enough.

[jarmak-nv]

Just to close the loop on this, I nuked one of my WSL Ubuntu 22.04 installs and started fresh following the page instructions and then doing a pip install of cuDF.

Our suggested test command of

 import cudf
 print(cudf.Series([1, 2, 3]))

So I'll remove the note when I add CUDA 12 on Monday because I think it could lead to extra confusion and it's not needed.

[bdice]

I proposed some changes. I'll approve to unblock this PR. Please resolve my suggestions as you see fit.

[jarmak-nv]

I like the changes - technical writing like this always takes two sets of eyes at a minimum so thank you!