Leo Fang
I don't think it's on the team's radar currently. Contribution is welcomed!
/ok to test
This is done as part of #597. Thanks, @vzhurba01!
tl;dr: No 😂
PyTorch depends on a ton of Python wheels. They also don't require a compiler to present at install time, and don't ship sdist themselves (see [this pypackaging page](https://pypackaging-native.github.io/key-issues/unexpected_fromsource_builds/) for reasons...
Non-blocking comments: - I discussed with @pciolkosz a while back about this, my opinion is that the entire `driver_api.h` should just be auto-generated. We can already do this same thing...
> I don't think the file to auto generate the header would contain less code, because we still want to adapt the API to our needs. It is about avoiding...
I'm all for this and would love to see it prioritized. This is a blocker for addressing #403. Based on my study (https://github.com/cupy/cupy/pull/7869), Thrust is extremely NVRTC unfriendly because [thrust/detail/type_traits.h](https://github.com/NVIDIA/cccl/blob/main/thrust/thrust/detail/type_traits.h)...
@vzhurba01 WDYT?
https://github.com/NVIDIA/cuda-python/discussions/388 could be another motivation.