CUDA in 22.05

[SomeoneSerge]

Issue description

This is a aggregate issue to track (the selected subset of) that state of CUDA in master before the 22-05 release.
It's concern is the world evaluated with { config.cudaSupport = true; config.allowUnfree = true; }

• The release will see the following changes that affect how we package and use cuda:

  • The new scope-based interface for consuming cuda-related packages, introduced together with the redistributable cuda packages in cudaPackages: overhaul of how we package cuda packages #167016
  • cudnn drops its propagatedBuildInputs which used to include cudatoolkit and cause a lot of obscure errors. Happens in: cudaPackages.cudnn: migrate to redist cuda, fix missing zlib #168748

• We currently have the following failures (we might not fix them before the release, but back-port later)

  • As of (at least) 2022-04-11 jax is failing on master, cf. jax: tests/test_sparse.py

    • test_sparse.py has been fixed upstream, but blocked in nixpkgs because of: python3Packages.jaxlibWithCuda build failure on x86_64-linux as of 891d0226 #169148

  • catboost

  • As of (at least) 2022-04-11, tensorflow-datasets is failing, cf. python3Packages.apache-beam build failure on x86_64-linux as of e29253ad #167445

  • xgboost

    CMake Error at CMakeLists.txt:140 (message):
    `BUILD_WITH_CUDA_CUB` should be set to `ON` for CUDA >= 11.4
    

    ...and now it's xgboost: checkPhase demands GPUs when built with cudaSupport #167931

  • python39Packages.mxnet
    CMake configurePhase fails

    Likely will be fixed in: cudatoolkit: 10.2 -> 11.5 #167250

  • cholmod-extra

    Tests fail: too large residuals

  • python3*Packages.chainer

    Complains about cupy version

    Marked as broken in: cudatoolkit: 10.2 -> 11.5 #167250

  • python310Packages.caffe

    CMake configurePhase fails

  • python39Packages.caffe

    Fixed by: fix pytorchWithCuda, fix cupy, upgrade cudnn #166784

  • blender: failing dependency (opensubdiv)

  • opensubdiv: fails at cmake configurePhase

    Likely will be fixed in: cudatoolkit: 10.2 -> 11.5 #167250

  • Pytorch

    Cmake error: python3Packages.pytorchWithCuda: uses unsupported cudatoolkit 10.1 #166403
    Build error: python3Packages.pytorchWithCuda build failure on x86_64-linux as of 3058e76a #161843
    Fixed by: fix pytorchWithCuda, fix cupy, upgrade cudnn #166784

cc @NixOS/cuda-maintainers

[samuela]

I think we may just need to mark some of these as broken. For example, chainer is requesting a truly ancient version of cupy: cupy>=7.7.0,<8.0.0. Latest stable cupy is v10.3.0.

cc'ing some maintainers:

• blender: @cillianderoiste @veprbl
• opensubdiv: @edolstra
• cholmod-extra: @jluttine
• python3Packages.chainer: @hyphon81
• python310Packages.caffe: @jb55

[SomeoneSerge]

The state of things at the time of the release landing:

I believe it would be still justifiable to backport some incremental improvements, fixes, and extensions as we make them. In particular, in the release we still

• haven't finished migrating downstream expressions to the redistributable cudaPackages, even the fundamental packages like
  • pytorch,
  • jax,
  • or tensorflow;
• don't have a good control of the closure size, e.g. when building docker images with deep learning packages: cudaPackages: clean "cudatoolkit" up #173462
• ..?