WIP: [DO NOT MERGE] introduce libcuml wheels

[jameslamb]

Replaces #6006, contributes to rapidsai/build-planning#33.

Proposes packaging libcuml as a wheel, which is then re-used by cuml-cu{11,12} wheels.

Blocked by rapidsai/cuvs#594

Notes for Reviewers

Benefits of these changes

• smaller wheels (see "Size Changes" below)
• faster compile times
  • no more re-compiling RAFT, thanks to introduce libraft wheels raft#2531
• less use of CI resources (only compiling once per CPU architecture / CUDA versions, instead of once per those + Python minor version)
• other benefits mentioned in Support dynamic linking between RAPIDS wheels build-planning#33

Wheel contents

libcuml:

• libcuml++.so (shared library) and its headers
• libcumlprims_mg.so (shared library) and its headers
• other vendored dependencies (CCCL, fmt)

cuml:

• cuml Python / Cython code and compiled Cython extensions

Dependency Flows

In short.... libcuml contains libcuml.so and libcumlprims_mg.so dynamic libraries and the headers to link against them.

• Anything that needs to link against cuML at build time pulls in libcugraph wheels as a build dependency.
• Anything that needs cuML's symbols at runtime pulls it in as a runtime dependency, and calls libcuml.load_library().

For more details and some flowcharts, see rapidsai/build-planning#33 (comment)

Size changes (CUDA 12, Python 3.12, x86_64)

wheel	num files (before)	num files (this PR)	size (before)	size (this PR)
libcuml	---	1766	---	289M
cuml	442	441	527M	9M
TOTAL	442	2207	527M	298M

NOTES: size = compressed, "before" = 2025-01-22 nightlies

how I calculated those (click me)

docker run \
    --rm \
    --network host \
    --env RAPIDS_NIGHTLY_DATE=2025-01-22 \
    --env CUML_NIGHTLY_SHA=01e19bba9821954b062a04fbf31d3522afa4b0b1 \
    --env CUML_PR="pull-request/6199" \
    --env CUML_PR_SHA="9d5100ec4589e20230a31817518427efa1e49c6d" \
    --env RAPIDS_PY_CUDA_SUFFIX=cu12 \
    --env WHEEL_DIR_BEFORE=/tmp/wheels-before \
    --env WHEEL_DIR_AFTER=/tmp/wheels-after \
    -it rapidsai/ci-wheel:cuda12.5.1-rockylinux8-py3.12 \
    bash

# --- nightly wheels --- #
mkdir -p ./wheels-before

export RAPIDS_BUILD_TYPE=branch
export RAPIDS_REF_NAME="branch-25.02"

# cuml
RAPIDS_PY_WHEEL_NAME="cuml_${RAPIDS_PY_CUDA_SUFFIX}" \
RAPIDS_REPOSITORY=rapidsai/cuml \
RAPIDS_SHA=${CUML_NIGHTLY_SHA} \
    rapids-download-wheels-from-s3 python ./wheels-before

# --- wheels from CI --- #
mkdir -p ./wheels-after

export RAPIDS_BUILD_TYPE="pull-request"

# libcuml
RAPIDS_PY_WHEEL_NAME="libcuml_${RAPIDS_PY_CUDA_SUFFIX}" \
RAPIDS_REPOSITORY=rapidsai/cuml \
RAPIDS_REF_NAME="${CUML_PR}" \
RAPIDS_SHA="${CUML_PR_SHA}" \
    rapids-download-wheels-from-s3 cpp ./wheels-after

# cuml
RAPIDS_PY_WHEEL_NAME="cuml_${RAPIDS_PY_CUDA_SUFFIX}" \
RAPIDS_REPOSITORY=rapidsai/cuml \
RAPIDS_REF_NAME="${CUML_PR}" \
RAPIDS_SHA="${CUML_PR_SHA}" \
    rapids-download-wheels-from-s3 python ./wheels-after

pip install pydistcheck
pydistcheck \
    --inspect \
    --select 'distro-too-large-compressed' \
    ./wheels-before/*.whl \
| grep -E '^checking|files: | compressed' \
> ./before.txt

# get more exact sizes
du -sh ./wheels-before/*

pydistcheck \
    --inspect \
    --select 'distro-too-large-compressed' \
    ./wheels-after/*.whl \
| grep -E '^checking|files: | compressed' \
> ./after.txt

# get more exact sizes
du -sh ./wheels-after/*

How I tested this

These other PRs:

• WIP: add libcuml Python builds devcontainers#442

[jameslamb]

/ok to test

[jameslamb]

Proposing bundling libcumlprims_mg.so + its headers into libcuml wheels:

• so libcuml++.so and the Cython extensions in cuml share 1 instance of this library
• to avoid needing to compile cumlprims_mg for both libcuml and cuml wheel builds

For conda packages, this isn't necessary because we distributed separate libcumlprims_mg conda packages that can be dynamically linked to.

We don't have the equivalent wheels and I don't think they'd be worth doing (given that they're only needed for cuML).

[bdice]

This plan seems fine to me. It reduces binary size and compile time, and provides the same behavior we already have for conda packages (except this is one wheel, rather than two conda packages).

[bdice]

This plan seems fine to me. It reduces binary size and compile time, and provides the same behavior we already have for conda packages (except this is one wheel, rather than two conda packages).

[bdice]

Is this a bug in cumlprims_mg's CMake?

[bdice]

Yes. Yes it is.

Let's wait for https://github.com/rapidsai/cumlprims_mg/pull/223 to merge, and then simplify this logic to use lib64/.