cuda-python icon indicating copy to clipboard operation
cuda-python copied to clipboard
verified publisher icon NVIDIA

Option to link statically against `libnvrtc_static.a`.

Open chr1sj0nes opened this issue 4 years ago • 3 comments

It would be great if it was possible to link against NVRTC statically.

chr1sj0nes avatar Nov 19 '21 14:11 chr1sj0nes

Sorry for lack of response @chr1sj0nes. Your ask implies you want to build cuda-python from source yourself. May I ask why?

leofang avatar Jun 12 '24 03:06 leofang

Within Google we build everything from source.

chr1sj0nes avatar Jun 13 '24 17:06 chr1sj0nes

Within Google we build everything from source.

Yes, so I was told 🙂 But, I was also told that Google maintains a fork of the whole universe internally (and patch them as needed). Since NVRTC APIs are dlopen'd, it should be possible to patch setup.py with appropriate flags in your internal fork.

Currently we're hesitant to support this feature because of two reasons:

  1. We rely on public PyPI to deliver software, and we must do everything we can to avoid generating over-sized wheels. Static linking simply does not play well with the constraints we're facing.
  2. We do not have enough bandwidth to develop/maintain/test this additional option (and because of 1 we would basically treat this option as nonexistent)

leofang avatar Jun 15 '24 02:06 leofang