adding wheel build for libcudf

[msarahan]

Description

Contributes to rapidsai/build-planning#33

Adds a standalone libcudf wheel, containing the libcudf C++ shared library.

Fixes #16588

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

Notes for Reviewers

Dependency Flows

My (@jameslamb)'s interpretation of the state we want to get to with this PR.

---
title: Build dependencies
---
flowchart TD
    A[libcudf] --> B[pylibcudf]
    A --> C[cudf]
    B --> C[cudf]
    B --> F[cudf-kafka]
    D[dask-cudf]
    E[cudf-polars]
    G[custreamz]

Loading

---
title: Runtime dependencies
---
flowchart TD
    A[libcudf] --> B[pylibcudf]
    B --> C[cudf]
    A --> C
    B --> E[cudf-polars]
    C --> D[dask-cudf]
    C --> F[cudf-kafka]
    C --> G[custreamz]
    F --> G

Loading

Size changes

wheel	size (before)	size (this PR)
libcudf.	---	440M
pylibcudf	469M	16M
cudf	478M	24M
cudf-polars	0.04M	0.04M
dask-cudf	0.05M	0.05M
TOTAL	947M	480M

NOTES: size = compressed, "before" = 2024-08-21 nightlies (58799d6)

how I calculated those (click me)

docker run \
    --rm \
    -v $(pwd):/opt/work:ro \
    -w /opt/work \
    --network host \
    --env RAPIDS_NIGHTLY_DATE=2024-08-21 \
    --env RAPIDS_NIGHTLY_SHA=58799d6 \
    --env RAPIDS_PR_NUMBER=15483 \
    --env RAPIDS_PY_CUDA_SUFFIX=cu12 \
    --env RAPIDS_REPOSITORY=rapidsai/cudf \
    --env WHEEL_DIR_BEFORE=/tmp/wheels-before \
    --env WHEEL_DIR_AFTER=/tmp/wheels-after \
    -it rapidsai/ci-wheel:cuda12.5.1-rockylinux8-py3.11 \
    bash

mkdir -p "${WHEEL_DIR_BEFORE}"
mkdir -p "${WHEEL_DIR_AFTER}"

cpp_projects=(
    libcudf
)
py_projects=(
    cudf
    pylibcudf
)
py_pure_projects=(
    cudf_polars
    dask_cudf
)

# TODO: calculate the date
for project in "${py_projects[@]}"; do
    # before
    RAPIDS_BUILD_TYPE=nightly \
    RAPIDS_PY_WHEEL_NAME="${project}_${RAPIDS_PY_CUDA_SUFFIX}" \
    RAPIDS_REF_NAME="branch-24.10" \
    RAPIDS_SHA=${RAPIDS_NIGHTLY_SHA} \
        rapids-download-wheels-from-s3 python "${WHEEL_DIR_BEFORE}"

    # after
    RAPIDS_BUILD_TYPE=pull-request \
    RAPIDS_PY_WHEEL_NAME="${project}_${RAPIDS_PY_CUDA_SUFFIX}" \
    RAPIDS_REF_NAME="pull-request/${RAPIDS_PR_NUMBER}" \
        rapids-download-wheels-from-s3 python "${WHEEL_DIR_AFTER}"
done


for project in "${py_pure_projects[@]}"; do
    # before
    RAPIDS_BUILD_TYPE=nightly \
    RAPIDS_PY_WHEEL_PURE="1" \
    RAPIDS_PY_WHEEL_NAME="${project}_${RAPIDS_PY_CUDA_SUFFIX}" \
    RAPIDS_REF_NAME="branch-24.10" \
    RAPIDS_SHA=${RAPIDS_NIGHTLY_SHA} \
        rapids-download-wheels-from-s3 python "${WHEEL_DIR_BEFORE}"
    
    # after
    RAPIDS_BUILD_TYPE=pull-request \
    RAPIDS_PY_WHEEL_PURE="1" \
    RAPIDS_PY_WHEEL_NAME="${project}_${RAPIDS_PY_CUDA_SUFFIX}" \
    RAPIDS_REF_NAME="pull-request/${RAPIDS_PR_NUMBER}" \
        rapids-download-wheels-from-s3 python "${WHEEL_DIR_AFTER}"
done

for project in "${cpp_projects[@]}"; do
    # before
    RAPIDS_BUILD_TYPE=nightly \
    RAPIDS_PY_WHEEL_PURE="1" \
    RAPIDS_PY_WHEEL_NAME="${project}_${RAPIDS_PY_CUDA_SUFFIX}" \
    RAPIDS_REF_NAME="branch-24.10" \
    RAPIDS_SHA=${RAPIDS_NIGHTLY_SHA} \
        rapids-download-wheels-from-s3 cpp "${WHEEL_DIR_BEFORE}"
    
    # after
    RAPIDS_BUILD_TYPE=pull-request \
    RAPIDS_PY_WHEEL_NAME="${project}_${RAPIDS_PY_CUDA_SUFFIX}" \
    RAPIDS_REF_NAME="pull-request/${RAPIDS_PR_NUMBER}" \
        rapids-download-wheels-from-s3 cpp "${WHEEL_DIR_AFTER}"
done

du -sh ${WHEEL_DIR_BEFORE}/*
du -sh ${WHEEL_DIR_BEFORE}
du -sh ${WHEEL_DIR_AFTER}/*
du -sh ${WHEEL_DIR_AFTER}

Related work

Corresponding devcontainers PR: rapidsai/devcontainers#271

We should merge that when this is just about ready, then re-run the failing devcontainers CI jobs here.

[msarahan]

Stuck on dlpack here. The failure is:

/__w/cudf/cudf/python/cudf/build/cp310-cp310-manylinux_2_28_aarch64/cudf/_lib/interop.cxx:1443:10: fatal error: dlpack/dlpack.h: No such file or directory
   1443 | #include "dlpack/dlpack.h"
        |          ^~~~~~~~~~~~~~~~~
  compilation terminated.

https://github.com/rapidsai/cudf/actions/runs/8662725105/job/23755706695?pr=15483#step:9:290

[msarahan]

@vyasr the build is passing now, but the tests seem to show that the library (libcudf.so) can't be found. It does seem like libcudf-cuXY is being installed OK, so I'm thinking it has to be some kind of RPATH issue. I don't see how you've addressed this in the RMM PR. Any advice?

I've inspected one of the libcudf wheels, and it has both lib and lib64 folders. The lib folder has nvcomp stuff, and lib64 has libcudf.so.

[vyasr]

You won't see this in the rmm PR because rmm is header-only, so there is no library there. The raft PR is a more useful example here. What you want to look at is actually the Python components of the libraft library, specifically the load.py module. The idea is that we never set RPATHs on libraries explicitly. Instead, we use ctypes to dlopen the library before we import any of the extension modules. This is a more robust solution than setting RPATHS because of the various ways in which it is possible for Python packages to coexist on a user's system (different environments, PYTHONPATH modification, etc).

[vyasr]

You have to set the variables before the rapids-cmake clone happens, which in this case is in the include(../../rapids_config.cmake) line. IOW the variables as currently set won't have any effect. Also, I think you'll need to specify the repo as vyasr/rapids-cmake, not just vyasr.

[bdice]

Error in the pip devcontainers:

  ERROR: Could not find a version that satisfies the requirement libcudf-cu12==24.10.*,>=0.0.0a0 (from versions: none)
  
  ERROR: No matching distribution found for libcudf-cu12==24.10.*,>=0.0.0a0

Is this blocked by rapidsai/devcontainers#271?

Edit: yes. From the description:

We should merge that when this is just about ready, then re-run the failing devcontainers CI jobs here.

[vyasr]

Almost there. Really only the build.sh change that's a blocker I think.

[vyasr]

Depending on merge order this can be removed.

[vyasr]

As above, this will be removable soon.

[jameslamb]

devcontainers job passed after merging rapidsai/devcontainers#271,

https://github.com/rapidsai/cudf/actions/runs/10515765176/job/29173481700?pr=15483

😁 thanks @AyodeAwe

Given that + all the approvals here, I'm gonna merge this. Thanks so much for all the help everyone!!!

[jameslamb]

/merge