cudnn: 8.7.0

[ConnorBaker]

Note
This is superseded by #215578.

Description of changes

Adds CUDNN 8.7.0.

Release notes are here: https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-870/release-notes/rel_8.html#rel-870.

Support matrix is here: https://docs.nvidia.com/deeplearning/cudnn/support-matrix/index.html#cudnn-cuda-hardware-versions.

Things done

• Built on platform(s)
  • x86_64-linux
  • aarch64-linux
  • x86_64-darwin
  • aarch64-darwin
• For non-Linux: Is sandbox = true set in nix.conf? (See Nix manual)
• Tested, as applicable:
  • NixOS test(s) (look inside nixos/tests)
  • and/or package tests
  • or, for functions and "core" functionality, tests in lib/tests or pkgs/test
  • made sure NixOS tests are linked to the relevant packages
• Tested compilation of all packages that depend on this change using nix-shell -p nixpkgs-review --run "nixpkgs-review rev HEAD". Note: all changes have to be committed, also see nixpkgs-review usage
• Tested basic functionality of all binary files (usually in ./result/bin/)
• 23.05 Release Notes (or backporting 22.11 Release notes)
  • (Package updates) Added a release notes entry if the change is major or breaking
  • (Module updates) Added a release notes entry if the change is significant
  • (Module addition) Added a release notes entry if adding a new NixOS module
  • (Release notes changes) Ran nixos/doc/manual/md-to-db.sh to update generated release notes
• Fits CONTRIBUTING.md.

[samuela]

@ConnorBaker OOC why did you close this?

[samuela]

btw there's now cuDNN v8.8.0 available. we should prob add that one. we need a new cuDNN version in order to upgrade CUDA 11.x -> 12.x

[ConnorBaker]

@samuela I've got a few PRs open and wasn't careful with my naming of branches and got confused between patch-1 .. patch-n. I'll make a new one for CUDNN 8.8.0 with a proper branch name.

I thought if I renamed the branch that GitHub would automatically update the PR to use that branch but I guess it just closed it.

[samuela]

no worries! a cuDNN 8.8.0 PR would be great!

[ConnorBaker]

Would it be okay to include 8.7.0 for CUDA 10.2? I don't recall what Nixpkgs wants to provide in terms of support for older versions (CUDNN 8.8.0 seems to have dropped support for CUDA 10.2).

[samuela]

I'm not against adding cuDNN 8.7.0 as well, but we will be wanting to cull old CUDA versions at some point. the maintenance burden is just too high. is there a reason you need CUDA 10.2 in particular?

[ConnorBaker]

I don’t need it personally. My thought was if we’re not removing it immediately I might as well update it to the last supported CUDNN. Does that make sense?

[SomeoneSerge]

On that note, I recall someone in the chat asking for cudnn 8.2.*. It's a dependency for onnxruntime. Maybe we should package that

• onnxruntime using nvidia docker hub images
• nvidia hub using cudnn 8.2.4

[samuela]

damn is there any way we can upgrade/patch onnxruntime around that?

[ConnorBaker]

With these unfree packages, is there any harm in providing multiple releases given that Hydra won’t cache them?

For example, if we wanted to provide all of the 8.x series for CUDNN, would that be problematic?

I would think just having them available would be helpful because it allows users to avoid having to vendor a specific release.

[SomeoneSerge]

damn is there any way we can upgrade/patch onnxruntime around that?

I haven't tried. But it's what they use in the trunk

I would think just having them available would be helpful

I think so too, especially that vendoring new releases out of nixpkgs tree is kind of cumbersome atm.

is there any harm in providing multiple releases

Only that it's extra attributes that can break and thus require keeping an eye on

[ConnorBaker]

It seems like 8.8.0 isn't available through the link we've been using previously: https://developer.download.nvidia.com/compute/redist/cudnn/

Is there anything to do but wait for it to appear there? Or is it possible they'd put it somewhere else?

For what it's worth, I was able to download it through the developer portal.

[ConnorBaker]

See #215578

[samuela]

Is there anything to do but wait for it to appear there? Or is it possible they'd put it somewhere else?

Hmm, well it does appear to be a recent release. It's also not showing up on https://developer.nvidia.com/rdp/cudnn-archive. If it doesn't show up in a week, ping me and we can reach out to them.