chore(dev): add pixi setup for cuda-bindings #1110
Conversation
|
Auto-sync is disabled for ready for review pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
cpcloud
force-pushed
the
pixi-for-cuda-bindings
branch
2 times, most recently
from
dc5b784 to
0583683
Compare
cuda_bindings/pyproject.toml
Outdated
| cu13 = { features = ["all", "test"], solve-group = "cu13" } | ||
|
|
||
| [tool.pixi.build-dependencies] | ||
| cuda-toolkit = ">=13,<14" |
There was a problem hiding this comment.
This pulls in a bunch of stuff we don't want. Can you reference the libraries in the all optional dependency group?
There was a problem hiding this comment.
I can't reference them in any sort automatic way, but does it matter since this is build-dependencies?
There was a problem hiding this comment.
The other thing is that including cuda-toolkit appears to set CUDA_HOME at build time, which I need to build cuda-bindings as part of environment activation.
There was a problem hiding this comment.
If there's a way to do that without including the kitchen sink, that'd be great!
There was a problem hiding this comment.
It matters for at least the cuda runtime library since we statically link that in
There was a problem hiding this comment.
Ah, ok. Then I'll have to figure out what's happening with CUDA_HOME and how that magically works with cuda-toolkit.
There was a problem hiding this comment.
Ok, I figured out the minimum set to build.
cpcloud
force-pushed
the
pixi-for-cuda-bindings
branch
from
0583683 to
39c8ff6
Compare
| cuda-bindings = { path = ".", editable = true } | ||
|
|
||
| [tool.pixi.environments] | ||
| cu13 = { features = ["all", "test"], solve-group = "cu13" } |
There was a problem hiding this comment.
Does the all feature end up grabbing wheels or does it map to the proper conda packages as defined in build-dependencies below?
There was a problem hiding this comment.
All of the pyproject.toml integration maps to pypi-dependencies: https://pixi.sh/dev/python/pyproject_toml/#dependency-section
There was a problem hiding this comment.
We can do it in a follow up, but we should ideally use https://pixi.sh/dev/reference/pixi_manifest/#conda-pypi-map-optional to map the pip wheels to the relevant conda packages in some kind of environment so we can support both wheel package based and conda package based environments.
cuda_bindings/pyproject.toml
Outdated
| cu13 = { features = ["all", "test"], solve-group = "cu13" } | ||
|
|
||
| [tool.pixi.build-dependencies] | ||
| cuda-cudart-dev = ">=13,<14" |
There was a problem hiding this comment.
In hindsight I think we need cuda-driver-dev since we look for cuda.h at build time as well
cpcloud
force-pushed
the
pixi-for-cuda-bindings
branch
3 times, most recently
from
a175181 to
b6eeebb
Compare
cpcloud
force-pushed
the
pixi-for-cuda-bindings
branch
from
b6eeebb to
df654df
Compare
| [tool.pixi.feature.all.dependencies] | ||
| libnvjitlink = "*" | ||
|
|
||
| [tool.pixi.build-dependencies] |
There was a problem hiding this comment.
We need to comprehensively capture these if we're adding conda package dependencies: https://github.com/conda-forge/cuda-python-feedstock/blob/69231e7f5de1c5a1314376f34308f82141f6112e/recipe/meta.yaml#L120-L131
In general we can't mix and match conda and pypi packages where it's kinda an all-or-nothing
2 participants
This PR adds a pixi environment for cuda-bindings