add CUDA 12.5 images

[jameslamb]

Contributes to rapidsai/build-planning#73

Proposes the following:

• adding CUDA 12.5 images
• PR builds:
  • +1 test job covering (cuda=12.5, arch=arm64)
• branch builds:
  • +1 test job covering (cuda=12.5, arch=x86_64)
  • -1 test job covering (cuda=12.2, arch=arm64)

context: rapidsai/build-planning#73 (comment)

Notes for Reviewers

Per offline discussion with @raydouglass , this would be accompanied by a deprecation notice in the RAPIDS 24.08 release stating that the CUDA 12.2 images will be removed in some future release (future release not yet determined).

[jakirkham]

Thanks James! 🙏

Am curious what you think about swapping where the x86_64 and aarch64 jobs are

For more context CUDA 12.2 added some interesting things like cuFile (for ARM) and HMM, which are helpful to test on CUDA 12.2 ARM specifically. Am not aware of a similar change in CUDA 12.5 (though could be missing it)

Given there are fewer ARM machines than x86_64, it may make sense to use x86_64 primarily in PRs and run ARM more on branches. That would mean less ARM jobs are needed during PR development

Anyways please let me know what you think 🙂

[jameslamb]

Thanks @jakirkham ! That's helpful context I didn't have. I just applied your suggestions.

[jakirkham]

Updating to pull in PR ( #690 ) now that is in branch-24.08

[jameslamb]

See this

9.818 Running dfg on cuspatial
9.939 Traceback (most recent call last):
9.939   File "/opt/conda/bin/rapids-dependency-file-generator", line 10, in <module>
9.939     sys.exit(main())
9.939              ^^^^^^
9.939   File "/opt/conda/lib/python3.11/site-packages/rapids_dependency_file_generator/_cli.py", line 121, in main
9.939     make_dependency_files(
9.939   File "/opt/conda/lib/python3.11/site-packages/rapids_dependency_file_generator/_rapids_dependency_file_generator.py", line 421, in make_dependency_files
9.940     raise ValueError(f"No matching matrix found in '{include}' for: {matrix_combo}")
9.940 ValueError: No matching matrix found in 'cuda_version' for: {'cuda': '12.5', 'arch': 'x86_64', 'py': '3.9'}

(build link)

This will be broken until rapidsai/cuspatial#1405 is merged. And that is blocked until CI in cuspatial is fixed (which could be done by merging rapidsai/cuspatial#1407).

[KyleFromNVIDIA]

Just one quick question. Approved so it doesn't block merging.

[jakirkham]

This need not be done in this PR. Just the discussion above ( #689 (comment) ) made me wonder

Should we do the same thing for the Linux version?

Suggested change

	ARG LINUX_DISTRO=ubuntu
	ARG LINUX_DISTRO_VER=22.04
	ARG LINUX_DISTRO=unset
	ARG LINUX_DISTRO_VER=unset

If this sounds reasonable, happy to send a 2nd PR with this change. Also ok not doing this if we don't think it makes sense

[jameslamb]

Yes in my opinion, this is a pattern that should be used in any Dockerfile where it's unacceptable to silently fall back to default values if configuration is not passed.

I just didn't want to go any further with it in this PR and grow the diff here.

I'd support a 2nd PR with this change.

[jakirkham]

One of the 11.8.0 tests failed on CI. Restarted it

That said, we are not changing 11.8.0 builds here and all other builds pass

Edit: Actually that won't work since the images get deleted after

[jakirkham]

Going to try restarting from scratch. If that doesn't work, we can discuss next steps on Monday

[raydouglass]

/merge

[jakirkham]

Thanks all! 🙏