Repository navigation
Make cuda-bindings buildable against CTK wheels? #692
Description
The meaning of the title is that pip install . from source can work without a local CTK environment (no need to set CUDA_PATH and all that). Everything is purely pip-wheel based.
This is actually doable, I just don't believe there is a serious use case because we really want users to not build from source, nor do I want to spend my time on enabling it, but I can be convinced. Anyway it's worth writing this up somewhere in the public.
The whole thing relies on the fact that building bindings just need a host compiler; no NVCC is needed. The steps is as follows:
- We list needed CTK wheels in the
[build-system.requires]session. They will provide the needed headers. - Since
setuptoolsdoes not provide a way for us to programmatically query the path to the isolated build environment at build time, we search the potential paths based on the assumption thatsys.pathmust be updated forsetuptoolsto find the build env. - We pass
<build env path>tosetuptoolsso that the CUDA headers can be found by gcc/msvc.
I enabled this for nvmath-python and it did worked like a charm, but the situation was a lot simpler than driver/runtime modules that we have here.
Activity
kkraus14 commentedon Jun 7, 2025
Sounds like a job for
path_finderif we split it off fromcuda.bindings😆leofang commentedon Jun 7, 2025
Oh, so you meant if we install
path_finderin the isolated build env, it knows the self location and can infer other build dependencies from this knowledge without guessing. Interesting!rparolin commentedon Jun 11, 2026
I'm going to close this issue due to inactivity. If we want to ultimately do this work please re-open the ticket.
leofang commentedon Aug 28, 2026
Took me a while to find this. Let's not close this. Keith and I discussed last week and we actually want this to happen.
leofang commentedon Aug 28, 2026
Closely related: #2060