Close CUDA 12 migration (redux)

[h-vetinari]

Since #5613 has been languishing for a while despite being essentially ready (aside from some open review requests to drop CUDA 12.4), let's actually get this over the finish line.

I made another clean-up push for the respective outstanding feedstocks. Below is the full list, but in any case, we already established that things are ready in #5613, so this just moves the needle even further into the "let's pull the trigger" area.

Closes #5390
Closes conda-forge/conda-forge.github.io#1963

Actively worked on

• Rebuild for CUDA 12 + PyTorch 2.4 aihwkit-gpu-feedstock#15
• Rebuild for CUDA 12 ambertools-feedstock#148
• Rebuild for CUDA 12 autoawq-feedstock#13
• Refresh CUDA 12 migrator cyrsoxs-feedstock#19
• faiss v1.8.0 faiss-split-feedstock#75
• Revive; rebuild for glog 0.7 & CUDA 12.0 graphvite-feedstock#6
• Rebuild for CUDA 12 w/arch + Windows support isce3-feedstock#64
• Rebuild for CUDA 12 w/arch + Windows support kaldi-feedstock#48
• Rebuild for CUDA 12 w/arch + Windows support kalpy-feedstock#24
• Rebuild for CUDA 12 w/arch + Windows support molgrid-feedstock#14
• Personal fork update to compile CUDA 12 seekr2_openmm_plugin-feedstock#21
• Rebuild for CUDA 12 tsnecuda-feedstock#12

Stagnant

• Rebuild for CUDA 12 w/arch + Windows support katago-feedstock#9
• Rebuild for CUDA 12 w/arch + Windows support lammps-feedstock#188
• Rebuild for CUDA 12 w/arch + Windows support meld-feedstock#23
• Rebuild for CUDA 12 w/arch + Windows support prismatic_split-feedstock#54
• [WIP] Refresh feedstock rmm-feedstock#31

Blocked upstream

• Rebuild for CUDA 12 w/arch + Windows support fairscale-feedstock#32 (not compatible with new pytorch)
• Rebuild for CUDA 12 w/arch + Windows support torchsparse-feedstock#30 (not compatible with new pytorch)

Dead feedstocks

• Rebuild for CUDA 12 w/arch + Windows support heavydb-ext-feedstock#65
• Rebuild for CUDA 12 w/arch + Windows support hdk-feedstock#56
• Rebuild for CUDA 12 w/arch + Windows support libtvm-feedstock#40
• https://github.com/conda-forge/timemory-feedstock
• https://github.com/conda-forge/tvm-py-feedstock (by virtue of depending on dead parent libtvm)

[conda-forge-webservices]

Hi! This is the friendly automated conda-forge-linting service.

I just wanted to let you know that I linted all conda-recipes in your PR (recipe/meta.yaml) and found it was in an excellent condition.

[h-vetinari]

Ping @jakirkham; I'd like to close this and then start discussing about a CUDA 12.4 or 12.6 migration (some packages already need newer versions than 12.0). I think we're ready here (and have been for quite a while), but since you invested so much time & energy on this migration I don't want to move forward without your input.

[jakirkham]

Hey Axel, sorry for the slow reply here

Axel and I discussed this offline after the last conda-forge meeting as we didn't get to this item in the meeting

Generally we agreed that the CUDA 12.0 migration has reached completion. There were few points I suggested we work on:

1. Try to get PRs close to completion through
   1. As CUDA 12 breaks the packaging format relative to CUDA 11 or earlier, thought we should try and get as many feedstocks on the other side of that break as possible
   2. Over the past few weeks Axel and I have made a best effort at this. The remainder are either unmaintained or will require substantial upstream work to adopt. So this is done
2. Clean up migrator vendoring needed to workaround bugs seen in pybind11 using feedstocks
   1. The combination of CUDA 12.0 + GCC 12 + pybind11 was problematic. We found rolling back to GCC 11 worked around these issues. So vendored migrators in feedstocks to use the older GCC
   2. Since then pybind11 included a fix in 2.12.0, which addressed this issue
   3. Have worked with maintainers to clean these up anywhere they remained. More context in comment: Close CUDA 11.8 migration #5340 (comment)
   4. So this is also done
3. Try to clear awaiting parents
   1. This isn't a hard blocker, but same as 1 we wanted to get this reasonably closed out due to the breaking nature of CUDA 12
   2. There are only 2 feedstocks without PRs. One appears unmaintained (like its parent). The other has an error as the feedstock is unmaintained and has fallen far behind
   3. In other words this is done

According to the status page 98% of PRs are made. Only 12 PRs remain open. With all of the work done above (thank you Axel for your hard work! 🙏), think we can safely declare this done

[jakirkham]

Thanks Axel! 🙏

[jakirkham]

Am noticing an issue with zip_keys: conda-forge/ucx-split-feedstock#195 (comment)

Axel, could you please take a quick look to make sure we are not missing anything here?

[h-vetinari]

I overlooked one of the places when taking over 1cf8857 where you had put in lower-case booleans, sorry.

Fixed in 23be5a0

[jakirkham]

No worries. Thanks Axel! 🙏