[RFC] Archive this project in favor of `cuda.core.(experimental).Linker`

[leofang]

As part of the greater RFC (NVIDIA/numba-cuda#128), on the pynvjitlink side I propose we do it in 3 steps

1. replace all custom CPython-based bindings by cuda.bindings.nvjitlink (almost done here: WIP: Switch to use new Python bindings from cuda-python/cuda-bindings  #111), and issue a deprecation warning
2. move all projects that need pynvjitlink today to either numba-cuda or cuda.core.Linker
   • the work of eliminating link= is done: Support linking code for device functions in declaration NVIDIA/numba-cuda#124, which made me believe numba-cuda can hide the linker entirely from the user
   • this is where all pynvjitlink-downstream projects need to chime in
3. after 1-2 release cycles, we archive pynvjitlink to free the RAPIDS team from the maintenance burden

[leofang]

Update: @bdice has kindly touched base with all teams that were using pynvjitlink and confirmed that with Step 2 wrapped up (and we skipped Step 1), we are now ready to proceed to Step 3 and archive this project. @bdice would you like to take the lead and finish the last mile? 🙂

[jameslamb]

@bdice would you like to take the lead and finish the last mile? 🙂

@leofang I can handle the mechanical parts of this RAPIDS (removing lingering references and build/publishing scripts, publishing a developer notice at https://docs.rapids.ai/notices/, etc.).

If you or @bdice could just help me with one part... could you give me a short statement to include in the docs and that notice helping people understand where to go to get the same functionality they were previously getting with pynvjitlink?

[bdice]

The end is in sight, but we're not quite done with Step 2 yet. We'll need to land NVIDIA/cudf#19604 and possibly some more follow-up work in RAPIDS (need to reassess after that PR lands), but I think there are no further requirements outside of RAPIDS.

As for docs and communicating what users should do now instead, here's an attempt at a deprecation message. @gmarkall @brandon-b-miller can you verify if this is complete and accurate?

The pynvjitlink library has been discontinued in favor of functionality in cuda.core.experimental.Linker. Users of numba-cuda>=0.16 can eliminate uses of from pynvjitlink import patch_numba_linker; patch_numba_linker(). The numba.cuda target will now automatically detect and enable nvjitlink when needed with no explicit configuration. All other public APIs of pynvjitlink are replaced by cuda.core and cuda.bindings functionality.

[gmarkall]

@bdice I think that's complete and accurate!