cudaPackages_12_2.cudatoolkit: init at 12.2.0

[Dessix]

Description of changes

Update cudaPackages: 12.1.1 -> 12.2.0

https://docs.nvidia.com/cuda/archive/12.1.1/ -> https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html

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/)
• 23.11 Release Notes (or backporting 23.05 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
• Fits CONTRIBUTING.md.

[SomeoneSerge]

Hi! This PR currently adds a run-file based cudaPackages_12_2.cudatoolkit (which we're trying to phase out) but not the redist packages like cudaPackages_12_2.cuda_nvcc.
We'd have to add the redist bit prior to merging at least because changing the default 12=12.1 to 12.2 effectively "removes" cudaPackages_12.cuda_nvcc that used to come from 12.0

(I'll follow up shortly with pointers on where to find the missing stuff)

[Dessix]

(I'll follow up shortly with pointers on where to find the missing stuff)

I believe I found some of it; I hadn't yet seen your comment and stumbled across the manifest files while browsing around the filesystem. I've added that, but I'm not sure what else may be missing.

For what it's worth, this PR is open to maintainer edits, so feel free to change whatever you need, if it's easier than commenting.

[SomeoneSerge]

Great! The manifests list sets of arguments to invoke build-cuda-redist-package.nix with, and you've already found the extension.nix where all the things get wired together

Check-list

• Verified the json:

❯ wget https://developer.download.nvidia.com/compute/cuda/redist/redistrib_12.2.0.json -O pkgs/development/compilers/cudatoolkit/redist/manifests/redistrib_12.2.0.json
❯ git diff
pkgs/development/compilers/cudatoolkit/redist/manifests/redistrib_12.2.0.json
│1151│}   │1151│}
\ No newline at end of file

An extra \n?

• Verifying the new package set:

❯ NIXPKGS_ALLOW_UNFREE=1 nix build --impure github:Dessix/nixpkgs/dev/dessix/update-cudatoolkit-12.2#cudaPackages_12_2.cuda_nvcc
error: hash mismatch in fixed-output derivation '/nix/store/yjk6l9qaaxnhi0ii7sims8bc7xa733nn-cuda_nvcc-linux-x86_64-12.2.91-archive.tar.xz.drv':
         specified: sha256-yJnFjQfbvjN6vq0BM3oq7HKn/UYAuVg4sEQzXIUZnGs=
            got:    sha256-1wOvCb6jmZtAyMBAlaeOx2zPArEQlAWTAKOpwr1S7GE=
error: 1 dependencies of derivation '/nix/store/mydj6kvjwj71i6snhyfsnnrrkw80qvh6-cuda_nvcc-12.2.91.drv' failed to build

Now that's odd.
CC @kmittman we get a hash mismatch between the manifest and the actual tars

[SomeoneSerge]

❯ wget https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/linux-x86_64/cuda_nvcc-linux-x86_64-12.2.91-archive.tar.xz
❯ sha256sum cuda_nvcc-linux-x86_64-12.2.91-archive.tar.xz
d703af09bea3999b40c8c04095a78ec76ccf02b11094059300a3a9c2bd52ec61  cuda_nvcc-linux-x86_64-12.2.91-archive.tar.xz
❯ cat redistrib_12.2.0.json | jq .cuda_nvcc.'"linux-x86_64"'.sha256
"c899c58d07dbbe337abead01337a2aec72a7fd4600b95838b044335c85199c6b"

[Dessix]

@SomeoneSerge Without the newline, the EditorConfig test gate failed. Should we solve this by fetch-ing manifests instead of checking them in?
As for the hash failures on manifest-referenced content, I have no idea.

[SomeoneSerge]

@Dessix we stage the manifests because this way we can use their contents in nix evaluation, without the "Import-From-Derivation" feature

Without the newline, the EditorConfig test gate failed

Oh, OK, let's just keep it

As for the hash failures on manifest-referenced content, I have no idea.

Let's wait for upstream to comment

[Dessix]

@SomeoneSerge

As for the hash failures on manifest-referenced content, I have no idea.

Let's wait for upstream to comment

Interestingly, the md5 for that file does match (b3b07d9b0b874c09a0cafadfc32b8760), but the sha256 does not. I wonder how that happened.

[kmittman]

Hi,
Thanks for catching this, I'm working on a fix ASAP.

[kmittman]

Apologies for the delay, resolved now https://developer.download.nvidia.com/compute/cuda/redist/redistrib_12.2.0.json
Left a comment about what happened in NVIDIA/build-system-archive-import-examples#8

[SomeoneSerge]

• Seems to build:

  ❯ NIXPKGS_ALLOW_UNFREE=1 nix build --impure github:Dessix/nixpkgs/dev/dessix/update-cudatoolkit-12.2#cudaPackages_12_2.cuda_nvcc
  ❯
  

• JSON matches what I download manually
• Commit message: previous commits seem to have followed the pattern, cudaPackages_12: init at 12.0.0, cudaPackages_12_1: init at 12.1.1; cudnn: don't break cudaPackages, etc

[Dessix]

I've split out the cudaPackages_12 remap to cudaPackages_12_2 for moving to a separate PR for potential ease of revert if it becomes necessary. The commit message now matches the pattern you described.

[SomeoneSerge]

Thank you @Dessix!

[ConnorBaker]

@Dessix can you reproduce these? I'm most concerned with the ones under Failures.

With ~/.config/nixpkgs/config.nix:

{
  allowUnfree = true;
  cudaSupport = true;
  cudaCapabilities = [ "8.9" ];
  cudaForwardCompat = false;
}

Running:

nix build --impure --keep-going -L .#legacyPackages.x86_64-linux.cudaPackages_12_2.{autoAddOpenGLRunpathHook,backendStdenv,cuda_cccl,cuda_cudart,cuda_cuobjdump,cuda_cupti,cuda_cuxxfilt,cuda_demo_suite,cuda_documentation,cuda_gdb,cuda_nsight,cuda_nvcc,cuda_nvdisasm,cuda_nvml_dev,cuda_nvprof,cuda_nvprune,cuda_nvrtc,cuda_nvtx,cuda_nvvp,cuda_opencl,cuda_profiler_api,cuda_sanitizer_api,cudatoolkit,cutensor,fabricmanager,libcublas,libcufft,libcufile,libcurand,libcusolver,libcusparse,libnpp,libnvidia_nscq,libnvjitlink,libnvjpeg,nccl,nsight_compute,nsight_systems,nvidia_fs}

I see the following (all but those under Failures are expected):

• Unavailable for x86_64-linux:
  • cuda_compat
  • libcudla
  • nsight_vse
  • visual_studio_integration
• Marked as broken
  • nvidia_driver
• Not updated for 12.2
  • cuda-library-samples
  • cuda-samples
  • tensorrt
• Failures
  • cuda_gdb: https://gist.github.com/ConnorBaker/9e52185252165424841bc8078dab2e0b
    • Missing libgmp dependency?
  • libcusolver: https://gist.github.com/ConnorBaker/229c5c17ed3a8189106deb0c67e63e7b
    • Missing libnvJitLink and libcusparse dependency?
  • libcusparse: https://gist.github.com/ConnorBaker/1065033fe0d45f9f30f11a25501f4520
    • Missing libnvJitLink dependency?
  • nsight_compute: https://gist.github.com/ConnorBaker/634cf3c033b2eab8b9dc2a265e2a3484
    • Looks like lots of missing dependencies?
  • nccl: https://gist.github.com/ConnorBaker/6cb69a17a8f125be882c3a4470f93c3c
    • Is NCCL not respecting the version of cudaPackages it should be built with? Notice in the log the reference to cuda_nvcc-11.8.89.
      • If that's the case, ignore this; it's a bug in how we package NCCL.

[Dessix]

@ConnorBaker All items mentioned above are failures present prior to this PR, for 12.0 and 12.1. They appear to be caused by the CUDA team's in-progress rework replacing cudatoolkit with the various single-subject packages, which seems to be incomplete based on the above.

I've reworked the PR slightly to un-vendor QT6, which will at least partially solve #239557, but several items (CollectX, neko) remain which may be targets for total removal.

This PR brings 12.2 to the same state as 12.1 and 12.0; further fixes are likely candidate targets for the CUDA team's larger-scoped rework, as I was unable to untangle the build-cuda-redist-package.nix relation (or lack thereof) with common.nix.

[ConnorBaker]

In that case, looks good to me! Thank you for taking a look at that, @Dessix.