11.7.1 had a huge ABI break

[kkraus14]

Solution to issue cannot be found in the documentation.

• I checked the documentation.

Issue

11.7.1 had a huge ABI break that moved from redefining all of the CUDA functions in Cython to externing them: NVIDIA/cuda-python@cfa118a

For any library that uses the Cython parts of cuda-python, this can lead to a runtime failure if something was built with 11.7.1 and then tries to run on 11.7.0, or vice versa because the function signature changes. I.E. if built with 11.7.1 and then tried to run on 11.7.0:

E   TypeError: C function cuda.ccudart.cudaStreamSynchronize has wrong signature (expected cudaError_t (cudaStream_t), got __pyx_t_4cuda_7ccudart_cudaError_t (__pyx_t_4cuda_7ccudart_cudaStream_t))

We should probably add a repodata patch to make anything pre 11.7.1 require <11.7.1 and anything post 11.7.1 require >=11.7.1,<12?

cc @jakirkham

The rmm packages in conda-forge are impacted by this as they're built with 11.7.1, but the rmm packages in the rapidsai channel are built with something older, presumably to avoid specifically this.

Installed packages

n/a

Environment info

n/a

[leofang]

@vzhurba01 @m3vaz This impacts CUDA Python wheel users too...

[bdice]

We may want to build rmm on conda-forge with the 11.7.0 release for consistency with the rest of RAPIDS… this is tough. @kkraus14 or others, do you have thoughts on that?

edit: this might be a bad idea but I’m not sure. I might pause that thought.

[leofang]

It should be doable. Only RMM 22.08.0 was built with CUDA Python 11.7.1. We could mark it as broken on conda-forge, rebuild with CUDA Python 11.7.0, and set a constraint on its runtime dependency to avoid 11.7.1, if everyone is OK. But it's better if we discuss rmm issues in https://github.com/conda-forge/rmm-feedstock, not here.

[leofang]

But it's better if we discuss rmm issues in https://github.com/conda-forge/rmm-feedstock, not here.

See conda-forge/rmm-feedstock#15.

[leofang]

@jakirkham Any thoughts on this? We definitely need a repodata patch, but AFAIK it only applies if cuda-python is pinned in the host or run section. If the build-time version is not specified (as in the case of rmm) repodata cannot help. Am I right?

[leofang]

Made two PRs to address this issue:

• Patch for cuda-python 11.7.1 conda-forge-repodata-patches-feedstock#325
• Pin cuda-python >=11.7.1,<12 #17

[jakirkham]

Have been out for a few weeks. So only now catching up on this.

Adding run_exports seems reasonable.

Some kind of repodata patch is needed. Commented there on how we might approach that.

[jakirkham]

AFAICT this has been addressed. Going to close this out. Please feel free to reopen if I've missed something. Thanks all! 🙏