Manually migrate for Python 3.12

[leofang]

For some reason the bot shows errors.

[leofang]

@conda-forge-admin, please rerender

[conda-forge-webservices]

Hi! This is the friendly automated conda-forge-webservice.

I just wanted to let you know that I started rerendering the recipe in #235.

[jakirkham]

Did a little reading. FWICT it looks like the bot was unable to solve the dependencies. Relevant snippet below:

  Could not solve for environment specs
  The following package could not be installed
  └─ cusparselt 0.2.0.1.* 1 does not exist (perhaps a typo or a missing channel).

[leofang]

@jakirkham @beckermr the PY312 migration status still says cupy is not solvable, despite I manually added the migrator (#235). It's unclear to the why the bot would complain about cusparselt. I checked on anaconda.org and saw the package is still there. Any clue if this is a disguise of other issues?

[beckermr]

Could be for sure. We should add a test case to the solver test package and figure out when we have time.

[leofang]

Thanks, Matt! Any tip for how I can unblock the migration for CuPy downstream for the time being?

[jakirkham]

Not to speak for Matt, but you could mark CuPy done for the Python 3.12 migration by making a PR similar to this one ( regro/cf-graph-countyfair#7 )

[leofang]

@conda-forge-admin, please rerender

[conda-forge-webservices]

Hi! This is the friendly automated conda-forge-webservice.

I just wanted to let you know that I started rerendering the recipe in #236.

[jakirkham]

Maybe this is relevant ( #238 )?

[jakirkham]

Unfortunately that didn't work, trying a different approach with PR ( #240 )

[leofang]

I'd love to get the bot bug fixed, but it seems we're still clueless. I had to disable the cusparselt support (#244) to make the bot happy. It finally works (migration is propagated to downstream: conda-forge/cuquantum-feedstock#54). Is there anything else we should do before closing this issue?

[jakirkham]

Think we should take a closer look at this list and see how we can cut that down

cupy-feedstock/recipe/meta.yaml

Lines 143 to 168 in 197bc44

	requirements:
	build:
	- {{ compiler("c") }}
	- {{ compiler("cxx") }}
	- {{ compiler("cuda") }}
	- sysroot_{{ target_platform }} 2.17 # [linux]
	- cross-python_{{ target_platform }} # [build_platform != target_platform]
	- python # [build_platform != target_platform]
	- cython >=0.29.22,<3 # [build_platform != target_platform]
	# TODO: clean up
	{% if cuda_major >= 12 %}
	- cuda-driver-dev # [build_platform != target_platform and linux]
	- cuda-cudart-dev # [build_platform != target_platform]
	- cuda-nvrtc-dev # [build_platform != target_platform]
	- cuda-nvtx-dev # [build_platform != target_platform]
	- cuda-profiler-api # [build_platform != target_platform]
	- libcublas-dev # [build_platform != target_platform]
	- libcufft-dev # [build_platform != target_platform]
	- libcurand-dev # [build_platform != target_platform]
	- libcusolver-dev # [build_platform != target_platform]
	- libcusparse-dev # [build_platform != target_platform]
	{% endif %}
	# optional dependencies for CUDA 11.2+
	- cudnn ~=8.8 # [build_platform != target_platform and not (aarch64 or (ppc64le and (cuda_compiler_version or "").startswith("12")))]
	- nccl ~=2.16 # [build_platform != target_platform]
	- cutensor ~=2.0 # [build_platform != target_platform]

Ideally we would get that down to build tools only. Atm there's quite a few libraries in there

Am sure there is some history I'm forgetting here (perhaps issues early on in CUDA 12 that have since been fixed?), but it may be worth revisiting

[leofang]

They are needed because of CuPy's autotool-like build system, which requires executable/linkable objects native to the build environment. Nothing has changed, and redundancy has been pruned.

[jakirkham]

Wanted to see what would happen (even if only to see errors)

xref: #246

[hmaarrfk]

this can be closed correct?

[jakirkham]

In terms of the feedstock being migrated for Python 3.12, yes that was done a long time ago

In terms of the bot errors, don't recall whether we were able to resolve those. We certainly tried lots of things that didn't end up working

[leofang]

Let's just close this since it is not going anywhere...