enforce wheel size limits, README formatting in CI

[jameslamb]

Description

Contributes to rapidsai/build-planning#110

Proposes adding 2 types of validation on wheels in CI, to ensure we continue to produce wheels that are suitable for PyPI.

• checks on wheel size (compressed),
  • to be sure they're under PyPI limits
  • and to prompt discussion on PRs that significantly increase wheel sizes
• checks on README formatting
  • to ensure they'll render properly as the PyPI project homepages
  • e.g. like how https://github.com/scikit-learn/scikit-learn/blob/main/README.rst becomes https://pypi.org/project/scikit-learn/

[dantegd]

Is this value coming from the existing wheel size, or coming from somewhere else?

[bdice]

This seems too big. cuML wheels appear to be closer to 550MB. Source: https://anaconda.org/rapidsai-wheels-nightly/cuml-cu12/files

Maybe set the threshold at 600MB.

[jameslamb]

This is the existing wheel size + a buffer. It varies by CPU architecture, Python version (because of Cython stuff), and CUDA version (because, for example, we don't use CUDA math lib wheels for CUDA 11).

The largest one I've seen clicking through logs on this PR was CUDA 11.8.0, Python 3.10, amd64:

checking 'final_dist/cuml_cu11-24.12.0a38-cp310-cp310-manylinux_2_28_x86_64.whl'
----- package inspection summary -----
file size
  * compressed size: 1.3G
  * uncompressed size: 2.2G
  * compression space saving: 42.2%
contents
  * directories: 72
  * files: 432 (86 compiled)
size by extension
  * .so - 2.2G (99.9%)
  * .py - 1.4M (0.1%)
  * .pyx - 1.2M (0.1%)
  * .0 - 0.2M (0.0%)
  * .ipynb - 0.1M (0.0%)
  * no-extension - 57.2K (0.0%)
  * .png - 51.3K (0.0%)
  * .pxd - 34.0K (0.0%)
  * .txt - 25.7K (0.0%)
  * .md - 10.3K (0.0%)
  * .h - 2.1K (0.0%)
  * .ini - 0.8K (0.0%)
largest files
  * (2.2G) cuml/libcuml++.so
  * (3.0M) cuml/experimental/fil/fil.cpython-310-x86_64-linux-gnu.so
  * (2.9M) cuml/fil/fil.cpython-310-x86_64-linux-gnu.so
  * (1.5M) cuml/cluster/hdbscan/hdbscan.cpython-310-x86_64-linux-gnu.so
  * (1.5M) cuml/svm/linear.cpython-310-x86_64-linux-gnu.so
------------ check results -----------
errors found while checking: 0

(build link)

So proposing setting this to around 200MB above that size, so we'd be notified if the binary size increased above that level.

There's nothing special about 1.5GB... it's already way way too big to be on PyPI. But proposing putting some limit so that we can get automated feedback from CI about binary size growth, and make informed decisions about whether to do something about it... similar to setting a coverage threshold for tests.

[bdice]

Aaahhh, but CUDA 11 is huge. We only did CUDA wheels work for CUDA 12. https://anaconda.org/rapidsai-wheels-nightly/cuml-cu11/files

[jameslamb]

Yeah exactly:

cuml/python/cuml/CMakeLists.txt

Lines 96 to 104 in 56e5e62

	# Link to the CUDA wheels with shared libraries for CUDA 12+
	if(CUDAToolkit_VERSION VERSION_GREATER_EQUAL 12.0)
	set(CUDA_STATIC_MATH_LIBRARIES OFF)
	else()
	if(USE_CUDA_MATH_WHEELS)
	message(FATAL_ERROR "Cannot use CUDA math wheels with CUDA < 12.0")
	endif()
	set(CUDA_STATIC_MATH_LIBRARIES ON)
	endif()

[bdice]

I’m indifferent as well. Let’s stick to the single definition for now.

[jameslamb]

alright sounds good, thanks for considering it. I'm glad these changes are helping to expose these differences and leading to these conversations 😊

[betatim]

I'd be a fan having two different limits. Mostly because for "not CUDA 11" the limit of 1.5GB might as well be "infinity". As in, if we ever reach it, it will be way to late to course correct.

Should I make a PR that uses Jams' suggestion for two limits?

[jameslamb]

@betatim sure! Go for it.

[jameslamb]

@betatim I've put up PRs in other repos following this suggestion, if you'd like something to copy from here in cuml:

• use different wheel-size thresholds based on CUDA version cuvs#469
• enforce wheel size limits, README formatting in CI cugraph#4754

[bdice]

Accepting the large threshold for now. Maybe we can make it depend on CUDA version? Or just wait until we drop CUDA 11.

[jameslamb]

/merge