libwholegraph wheels: use nvidia-nccl wheels instead of vendoring libnccl.so by jameslamb · Pull Request #284 · rapidsai/cugraph-gnn

jameslamb · 2025-08-25T18:52:32Z

Fixes #281

Similar to rapidsai/cuvs#827, proposes that wheels get their copy of libnccl.so at runtime from nvidia-nccl-cu{12,13} wheels, instead of vendoring a copying inside libwholegraph.

Notes for Reviewers

Benefits of these changes

I think this fill fix the nightly tests
- @alexbarghi-nv investigated offline and together with some other people found that "competing libnccl between this project and pytorch" may be to blame for the failing nightly tests
smaller libwholegraph wheels
- 120 MiB smaller!
- now they can go on pypi.org without needing an exception 🎉 🎉 🎉
smaller disk footprint for environments
- via avoiding multiple copies of libnccl.so
reduced risks of runtime bugs caused by multiple competing libnccl.so being loaded

…nccl.so

copy-pr-bot · 2025-08-25T18:52:35Z

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

jameslamb · 2025-08-25T18:52:40Z

/ok to test

jameslamb · 2025-08-25T19:32:08Z

I think this is ready for review!

cugraph-pyg wheel tests are failing like this:

/pyenv/versions/3.10.18/lib/python3.10/site-packages/pylibcudf/__init__.py:13: in <module>
    from . import (
E   ImportError: libcudf.so: cannot open shared object file: No such file or director

(build link)

But probably because of the ongoing libcudf.so-loading issues that I think should be fixed by rapidsai/cudf#19786

jameslamb · 2025-08-25T19:36:20Z

ci/build_wheel_pylibwholegraph.sh

 echo "libwholegraph-${RAPIDS_PY_CUDA_SUFFIX} @ file://$(echo "${LIBWHOLEGRAPH_WHEELHOUSE}"/libwholegraph_*.whl)" >> "${PIP_CONSTRAINT}"

-export SKBUILD_CMAKE_ARGS="-DBUILD_SHARED_LIBS=ON;-DCMAKE_MESSAGE_LOG_LEVEL=VERBOSE;-DCUDA_STATIC_RUNTIME=ON;-DWHOLEGRAPH_BUILD_WHEELS=ON"
+export SKBUILD_CMAKE_ARGS="-DBUILD_SHARED_LIBS=ON;-DCMAKE_MESSAGE_LOG_LEVEL=VERBOSE;-DCUDA_STATIC_RUNTIME=ON"


Not strictly related, but noticed while I was looking at logs to confirm the NCCL stuff was working correctly.

CMake Warning: Manually-specified variables were not used by the project: CUDA_STATIC_RUNTIME WHOLEGRAPH_BUILD_WHEELS

(build link)

Think we still want CUDA_STATIC_RUNTIME because it could get passed through here:

cugraph-gnn/python/pylibwholegraph/CMakeLists.txt

Line 38 in e1ec288

add_subdirectory(pylibwholegraph/binding)

But WHOLEGRAPH_BUILD_WHEELS doesn't do anything anywhere in this project.

git grep '_BUILD_WHEEL'

LGTM, thanks!

jameslamb · 2025-08-25T19:37:09Z

/ok to test

jameslamb · 2025-08-25T20:37:33Z

This will need an admin merge because of the failing check-nightly-ci test...I can do that once this has approvals.

alexbarghi-nv

👍

alexbarghi-nv · 2025-08-25T21:37:41Z

@jameslamb you have my permission to admin merge once we have packaging/CI approval

jameslamb · 2025-08-25T21:38:41Z

Thank you!

gforsyth

Very nice improvement, let's gooooo!

gforsyth · 2025-08-26T20:19:55Z

python/cugraph-pyg/pyproject.toml


 # PyPI limit is 100 MiB, fail CI before we get too close to that
-max_allowed_size_compressed = '75M'
+max_allowed_size_compressed = '10Mi'


ooooh yeaahhh

gforsyth · 2025-08-26T20:20:23Z

python/libwholegraph/pyproject.toml

+# PyPI limit is 100 MiB, fail CI before we get too close to that
+max_allowed_size_compressed = '80Mi'


Even better!

jameslamb · 2025-08-26T20:24:54Z

Thanks @gforsyth ! Ok I'm going to admin-merge this and then manually trigger the nightly tests.

jameslamb · 2025-08-26T22:34:19Z

All of the wheel tests passed on the manually-triggered nightly run!

https://github.com/rapidsai/cugraph-gnn/actions/runs/17250350223

Just one conda test failed, and it looks to me like something temporary while downloading a large dataset. I just restarted it... I'm hopeful it'll pass and that nightlies will be working here again.

#284 devendored the `libnccl.so` from `libwholegraph` wheels. That wasn't extended to cover CUDA 13 wheels when added CUDA 13 support, so `libwholegraph-cu13` wheels have been depending on getting this transitively from `raft` and `cugraph` at runtime 😬 This fixes that. Authors: - James Lamb (https://github.com/jameslamb) Approvers: - Bradley Dice (https://github.com/bdice) URL: #428

libwholegraph wheels: use nvidia-nccl wheels instead of vendoring lib…

0234047

…nccl.so

jameslamb added improvement Improves an existing functionality non-breaking Introduces a non-breaking change labels Aug 25, 2025

jameslamb changed the title ~~WIP: libwholegraph wheels: use nvidia-nccl wheels instead of vendoring libnccl.so~~ libwholegraph wheels: use nvidia-nccl wheels instead of vendoring libnccl.so Aug 25, 2025

remove unused flag, reduce size thresholds

ce5f97b

jameslamb commented Aug 25, 2025

View reviewed changes

jameslamb requested review from KyleFromNVIDIA and alexbarghi-nv August 25, 2025 20:36

jameslamb marked this pull request as ready for review August 25, 2025 20:37

jameslamb requested review from a team as code owners August 25, 2025 20:37

alexbarghi-nv approved these changes Aug 25, 2025

View reviewed changes

jameslamb mentioned this pull request Aug 26, 2025

Update rapids-dependency-file-generator #285

Merged

gforsyth approved these changes Aug 26, 2025

View reviewed changes

jameslamb merged commit be71c89 into rapidsai:branch-25.10 Aug 26, 2025
48 of 50 checks passed

jameslamb deleted the devendor-libnccl branch August 26, 2025 20:25

jameslamb mentioned this pull request Mar 10, 2026

libwholegraph: declare nvidia-nccl dependency for CUDA 13 wheels #428

Merged

		# PyPI limit is 100 MiB, fail CI before we get too close to that
		max_allowed_size_compressed = '80Mi'

Conversation

jameslamb commented Aug 25, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Notes for Reviewers

Benefits of these changes

Uh oh!

copy-pr-bot bot commented Aug 25, 2025

Uh oh!

jameslamb commented Aug 25, 2025

Uh oh!

jameslamb commented Aug 25, 2025

Uh oh!

jameslamb Aug 25, 2025

Choose a reason for hiding this comment

Uh oh!

alexbarghi-nv Aug 25, 2025

Choose a reason for hiding this comment

Uh oh!

linhu-nv Aug 26, 2025

Choose a reason for hiding this comment

Uh oh!

jameslamb commented Aug 25, 2025

Uh oh!

jameslamb commented Aug 25, 2025

Uh oh!

alexbarghi-nv left a comment

Choose a reason for hiding this comment

Uh oh!

alexbarghi-nv commented Aug 25, 2025

Uh oh!

jameslamb commented Aug 25, 2025

Uh oh!

gforsyth left a comment

Choose a reason for hiding this comment

Uh oh!

gforsyth Aug 26, 2025

Choose a reason for hiding this comment

Uh oh!

jameslamb Aug 26, 2025

Choose a reason for hiding this comment

Uh oh!

gforsyth Aug 26, 2025

Choose a reason for hiding this comment

Uh oh!

jameslamb Aug 26, 2025

Choose a reason for hiding this comment

Uh oh!

jameslamb commented Aug 26, 2025

Uh oh!

Uh oh!

jameslamb commented Aug 26, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

jameslamb commented Aug 25, 2025 •

edited

Loading