cudaPackages.cudatoolkit: 12.3.0 -> 12.3.2; init 12.4.1

cudaPackages.cudnn: update for 9.x

cudaPackages.cuda_cudart: remove stubs workaround

python3Packages.openai-triton: update substituteInPlace and cudaPackages outputs

individual allowFHSReferences, don't use getOutput, refactor setupCudaHook

fixups

fixup

fixup

switch back to cudaFlags; various simplifications

fixup eval errors

remove extra whitespace

fixups

fixup

fixup

fixups

fixups

fixups

fixups

fixups

fixups

fixups

fixups

fixups

fixups

fixups

fixups

fixup

fixup

fixup

mapAttrs mkOption to avoid forgetting to apply it to an argument

alphabetize

Add latest from cuda-redist-find-features; must update to reflect nesting of index in data.indices

update for data.indices and file shuffle

wip

wip

brain hurt ;(

CUDA packages should be visible on all platforms, but not necessarily buildable

switch to utils.mkMissingPackagesBrokenConditions pattern to avoid missing attribute errors

cudaPackages now evaluates deeply

wip

wip

wip

use lazyAttrsOf to avoid strictly evaluating during checks

wip

overrideAttrs doesn't deeply merge attribute sets

fix missing src for cudnn_samples

cudnn_samples: mark as broken for now

cudaPackages: add major versions to avoid need sort and filter packages

cudaPackages.cudnn: fix selection logic (was not using result of filterAttrs)

cudaPackages.cuquantum: fixup

cudaPackages.nccl: broken on CUDA pre-11.4 due to API changes

cudaPackages: add note about recurseIntoAttrs and nixpkgs-review

cudaPackages: temporarily add recurseIntoAttrs on all package sets

cudaPackages.utils: fix usages of mkMissingPackagesBrokenConditions

cudaPackages.tensorrt: use cudatoolkit prior to CUDA 11.4

cudaPackages: set name expressly to make clear where packages originate

cudaPackages.cudnn: requires CUDA 10.1 or later

cudaPackages.libnvvm_samples: allow FHS references

cudaPackages.nsight_compute: fix

cudaPackages.libcal: correct dependency on ucc

cudaPackages.libcublasmp: correct selection logic

cudaPackages: deep overriding of pkgs

cudaPackages: re-add reference to self to fix package selection

Prior to this change, packages like CUDNN would take the newest version of packages like cudatoolkit (which is to say
packages created through packagesFromDirectoryRecursive). That is, cudaPackages_11_0.cudnn_8_5.buildInputs would
reference cudaPackages.cudatoolkit rather than cudatoolkit from the same cudaPackages.

!fixup 3724a77

wip

cudaPackages.setupCudaHook: wip

cudaPackages_12_5: fixes

cudaPackages.cudnn_samples: require cudnn be available

cudaPackages: guard invalid eval behind badPlatforms

cudaPackages.tensorrt: don't rely on passthru for buildInputs

While this may complicate overriding the package, it does resolve an infinite recursion error.

cudaPackages: create dummy outputs to avoid missing attribute errors

cudaPackages.libcusparse_lt: add overrides and constraints

cudaPackages.tensorrt: handle constraints for unsupported architectures

cudaPackages.cuda_compat: allow FHS references

cudaPackages: fix src selection logic

TODO: update

cudaPackages: enable override and implement per-output overrides

cudaPackages: add setup hook for cuda_cudart.stubs to add to NIX_LDFLAGS

cudaPackages.cuda_cudart.stubs: set LD_LIBRARY_PATH in setup hook

cudaPackages: move setup-cuda-stubs-hook to manifest-builder

wip

fixup after rebase on 2024-07-10

regen JSON as of cuda-redist-find-features@239ee25f2ed08a27e24762be384361d17b8911b8

Thanks to the work of SomeoneSerge, the lib.get* family of output selectors now work!

In order to keep closure sizes small, the `dev` output propagates only `bin`, `include`, and
`lib` outputs. The `include` output is new -- it is the `include` directory, and was separated
out from the `dev` output to allow for more granular dependencies between derivations. As an example,
cuda_cudart.include depends on cuda_nvcc.include and cuda_cccl.include, but adding a dependency on
the dev output of either would greatly increase the size of the closure relative to the tiny header files.

Since `dev` is the default output selected by `make-derivation.nix`, in general the `lib.get*` family of
functions should only be necessary for retrieving an output which contains a specific directory. This happens
frequently when trying to slim down a closure (retrieving only the `include` output) or because you need the
full path to a directory (e.g., `(lib.getLib cuda_cudart) + "/lib"`).

Caffe is unmaintained in and out of Nixpkgs.

…verrideAttrs

`cuda_cccl.dev` is now propagated by `cuda_cudart.dev`, which is provided by `cuda_cudart`.

`opencv.cxxdev` provides all necessary CUDA utilities for this package.