The nvcc packages do not honor cuda-version

[leofang]

Solution to issue cannot be found in the documentation.

• I checked the documentation.

Issue

@pentschev reported offline that nvcc_linux-64 11.8 can coexist with a CUDA 12.x environment. It turns out that the recipe here does not make cuda-version a dependency:

channels:
  - rapidsai-nightly
  - dask/label/dev
  - conda-forge
  - nvidia
dependencies:
# Base
  - python=3.10
  # - cudatoolkit=11.8
  - cuda-version=12.5
...
  - nvcc_linux-64=11.8
...

Installed packages

n/a

Environment info

n/a

[leofang]

Based on the conclusion here, it seems the preference would be to either fix the repodata or add nvcc wrapper to cuda-version's run_constrained:
https://github.com/conda-forge/cuda-version-feedstock/blob/c8e4d1cc2c11b78ca53cff1cd2978afaab3d3651/recipe/meta.yaml#L14-L18

[jakirkham]

In CUDA 11 (and earlier), constraining of CUDA version and installing runtime libraries was done by the same package cudatoolkit (as outlined in issue: conda-forge/conda-forge.github.io#1963 )

nvcc does add cudatoolkit to run_exports, but this only influences Conda packages that depend on nvcc during their build

nvcc-feedstock/recipe/meta.yaml

Lines 27 to 30 in 5539c2f

	run_exports:
	strong:
	- cudatoolkit >={{ cuda_compiler_version }},<{{ cuda_major + 1 }}
	- sysroot_{{ cross_target_platform }} >={{ c_stdlib_version }} # [linux]

A natural extension of this for users installing nvcc in their development environments would be to add cudatoolkit as a runtime dependency

However it appears nvcc currently does not have cudatoolkit in requirements/run, which differs from nvcc's other run_exports like sysroot that is in requirements/run

nvcc-feedstock/recipe/meta.yaml

Lines 40 to 42 in 5539c2f

	run:
	- sed # [linux]
	- sysroot_{{ cross_target_platform }} >={{ c_stdlib_version }} # [linux]

As noted above, cuda-version already constrains cudatoolkit to match

So a reasonable fix would simply be to add cudatoolkit to requirements/run with an appropriate version constraint (likely just copying what run_exports already has)

[bdice]

I considered some other options but I think @jakirkham's proposal above is the cleanest. It avoids pushing any complexity on the conda solver for users of CUDA 12, which is ideal. We know the cuda-version solve path is rather difficult for CUDA 12 and we don't want legacy (CUDA 11) constraints to negatively impact the solver going forward.

[bdice]

I opened #109 with @jakirkham's proposed fix. We may need a repodata patch for older versions, too?

[jakirkham]

Agreed we will want a repodata patch for older versions

[jakirkham]

@mtjrider to look into making a repodata patch

Here is a good example. The patch would be to nvcc using the change in PR: #109