[FEA]: support system ctk install for python package

[ZzEeKkAa]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this request and that I agree to the Code of Conduct

Area

Not sure

Is your feature request related to a problem? Please describe.

I was recently building docker image where I need to use ctk both at C++ level and python level. At python level I was also using cuda.cccl and current docs does not provide a way to reuse system install ctk, only packages installed through pypi. That results into duplicate package install and blows container size.

Describe the solution you'd like

Support for system install packages. At nvmath python we have [sysctk1*] subpackage support that installs all the requirements, but does not install pypi ctk. https://github.com/NVIDIA/cuda-python suppose to work with system install.

Describe alternatives you've considered

I was able to hack system ctk support together by using this:

pypi dependencies:

sysctk13 = [
  "cuda-bindings>=13.0.0,<14.0.0",
  "numba-cuda>=0.23.0",
]

with

apt install -y cuda-nvrtc-13-1 libnvjitlink-13-1 cuda-cudart-13-1 cuda-nvcc-13-1 libnvvm-13-1

and setting CUDA_PATH=/usr/local/cuda-13.1 to help cuda.pathfinder find headers from system install.

based on the

cu13 = [
  "cuda-bindings>=13.0.0,<14.0.0",
  "cuda-toolkit[nvrtc,nvjitlink,cudart,nvcc,nvvm]==13.*",
  "numba-cuda[cu13]>=0.23.0",
]

Additional context

Nvmath reference:
https://github.com/NVIDIA/nvmath-python/blob/main/pyproject.toml#L87-L94

sysctk12 = [
    "cuda-bindings >=12.9.2,<13",
    "cutensor-cu12 >=2.3.1",
]
sysctk13 = [
    "cuda-bindings >=13.0.1,<14",
    "cutensor-cu13 >=2.3.1",
]
# ...
sysctk12-dx = [
    "nvmath-python[sysctk12]",
    "nvmath-python[dx]",
]
sysctk13-dx = [
    "nvmath-python[sysctk13]",
    "nvmath-python[dx]",
]

Note: test-cu1* optional dependencies may be moved to [dependency-groups] to avoid distributing them with wheels.

[github-actions]

Hi @ZzEeKkAa!

Thanks for submitting this issue - the CCCL team has been notified and we'll get back to you as soon as we can!
In the mean time, feel free to add any relevant information to this issue.

[jrhemstad]

@shwina can correct me if I'm wrong, but I don't believe this will be possible because cuda.compute depends on updates to the underlying CCCL C++ headers and will not be guaranteed to work with whatever headers exist in the CTK install.

That results into duplicate package install and blows container size.

This is surprising to hear. The CCCL C++ headers package should be trivially small given they are just purely headers and don't contain any binaries.

[shwina]

I think @ZzEeKkAa's concern is more about the size of the dependencies like nvJitLink, nvrtc etc.,

We'll always use our own headers (we ship with them), but I see @ZzEeKkAa's point that it would be nice to use system CUDA libraries if available.

My only question/concern is around the convention pip install cuda-packagename[sysctk12 | sysctk13] to skip installing CUDA libraries. Is that a convention we want to agree upon across our different Python offerings? (cc: @leofang). The consistency would be good because it then allows expressing dependencies like so:

sysctk12 = [
    "cuda-bindings >=12.9.2,<13",
    "cutensor[sysctk12]",
    "cuda-cccl[sysctk12]",
    ...etc
]