cudaPackages: bump cudaPackages_11 -> cudaPackages_12

[SomeoneSerge]

CC @NixOS/cuda-maintainers

Blockers:

• pytorch official support for cuda 12: [CUDA][CUDA 12] CUDA 12 Support Tracking Issue pytorch/pytorch#91122
• ...

Thanks @ConnorBaker for the pytorch link

[nviets]

Will this bump happen after #220341? I am attempting to build 11_8 now on your latest branch to see if it fixes #224150. No idea if xgboost will be happy with 12, but I'll give it a shot. Thanks @SomeoneSerge for keeping me in the loop!

[SomeoneSerge]

Will this bump happen after #220341?

I guess the bump will happen as soon as someone opens the bump PR and runs a nixpkgs-review on it (with --extra-nixpkgs-config { cudaSupport = true; }, which is going to be very compute-intensive). Right now the concern is that pytorch doesn't yet support CUDA12, but there are probably others, which is why the safest way to see whether we can update is by trial.

There's no reason we couldn't first update to 11.8: that would likely cause fewer regressions than 12, and would enable you to fix xgboost without resorting to overrides. Still, we'd have to run the builds first

[nviets]

Thanks, xgboost 1.7.5 should be safe once I get it building with 11_8 because I added the version argument to options.

[samuela]

Upgrading to CUDA 12 will happen eventually, but I don't realistically see it happening for another few weeks at least.

[SomeoneSerge]

Let's open a bump to 11.8 then. @nviets would you like to take that on?

[nviets]

Sure - what does the bump entail? An update to the default in all-packages.nix and a nixpkgs-review to see if anything downstream breaks?

[nviets]

I am not familiar with the derivation for pytorch, but I see it's coded to accept nixpkgs' default library for CUDA. Should that be left as-is to build with 11.8, or should it be an option? I was under the impression in my xgboost update that it's better to have a configurable version of CUDA using cudaPackages ? {}. If pytorch should also have the option, should it be handled here or in a separate PR?

[SomeoneSerge]

We should've bumped long time ago, now we need to do it for #259068

[samuela]

Agree that we should bump, but why is it needed for #259068?

[SomeoneSerge]

Agree that we should bump, but why is it needed for #259068?

To avoid spawning extra cudaPackageses: https://github.com/NixOS/nixpkgs/pull/259068/files#diff-7c765cc9768cdae049f8712b4ef719bc94ec8ec2e8df2dd50ddf14c6acd4e580R13927

[samuela]

ah right, my own comment all along :P

[Turakar]

Hello, what is the state of this? I am currently using dnf on Fedora 37 for all CUDA stuff, but I encounter issues with the upgrade to 38 due to GCC incompatibilities with CUDA on Fedora 38. This led me to maybe try CUDA with Nix, but it seems like we still do not have any CUDA 12 support, although Torch moved to CUDA 12.1 as the new default?

[SomeoneSerge]

This led me to maybe try CUDA with Nix, but it seems like we still do not have any CUDA 12 support, although Torch moved to CUDA 12.1 as the new default?

Hi! Note that cudaPackages is just the default we set, and you can override* it locally. That said, it's definitely the time we bumped the default version, we were just caught up with other stuff

*An example of how to override the default (do ping matrix or discourse for details if needed!):

import <nixpkgs> {
  config.allowUnfree = true;
  overlays = [
    (final: prev: { cudaPackages = final.cudaPackages_12; })
  ];
}`

[Turakar]

Hi! Note that cudaPackages is just the default we set, and you can override* it locally. That said, it's definitely the time we bumped the default version, we were just caught up with other stuff

Thanks, your suggestion worked! It seems like cuDNN 8.9.2 is not packaged, yet. I think it might be easier for me to go for CUDA 11.8 for now. It is in unstable. Nice 🙂

[SomeoneSerge]

@Turakar note the linked PR too: I had a few evaluation errors when I changed the default

[Dessix]

It is in unstable. Nice 🙂

I just made a PR to add cudnn 8.9.6.50 and get it supported in 12.2, if that helps at all: #271201