Make `pixi-build-python` recognize `cuda` compiler

[leofang]

Right now it does not and we have to work around the resolver errors like this:

[package.build.config]
# For some reason it has to be "cuda-nvcc" instead of "cuda" as documented by pixi-build
compilers = ["c", "cxx", "cuda-nvcc"]

[package.build-dependencies]
# For some reason "cuda-nvcc" in config.compilers is not enough
cuda-compiler = "*"

But this technically is not correct (from the perspective of conda recipes).

[leofang]

cc @cpcloud for vis

[ruben-arts]

I believe we added it to replace this line:

https://github.com/conda-forge/openmpi-feedstock/blob/a6741814b591b9e4c5322d6041b102bc41e060d0/recipe/recipe.yaml#L46-L48

What would the cuda-nvcc mean?

[cpcloud]

@ruben-arts It seems like cuda-nvcc (ideally just cuda) should bring in the appropriate cuda compiler analogous to how this works with the other compilers.

[ruben-arts]

Ah I think we missed to add the default compiler for cuda to the default_compilers

You can pass the variants using the build-variants field in the pixi.toml to add this.

That would look something like:

[workspace.build-variants]
cuda_compiler = ["cuda-nvcc"]

That said, it would be great if you could instruct us what would be best here. The current default compiler code is located here:

pixi-build-backends/crates/pixi-build-backend/src/compilers.rs

Line 34 in e6cf8b2

pub fn default_compiler(platform: &Platform, language: &str) -> String {

[cpcloud]

I would guess that the right default compiler for cuda would be nvcc, but @leofang would know for sure.

[leofang]

If default_compiler is mapping the languages to the user-facing conda compiler packages, as defined in the global pinning, then cuda-nvcc is the right package to use for the cuda language on Linux/Windows.

The (now archived) nvcc package was used in CUDA 11 and before. We shifted to the "new layout" starting CUDA 12.0 and cuda-nvcc was introduced as part of this effort. IIRC somewhere in this doc the history of old/new CUDA packages is explained.

[cpcloud]

Ah, okay. I didn't realize there was anything called nvcc, thanks for clarifying!

[baszalmstra]

I think we can add a default variant for cuda_compiler. Then things would work out of the box.