Remove cudf components by vyasr · Pull Request #1083 · rapidsai/rapidsmpf

vyasr · 2026-06-05T17:11:54Z

This PR removes all the parts of rapidsmpf that rely on libcudf from rapidsmpf. Those parts have all been moved to the cudf_streaming library in the cudf repository.

Tests and benchmarks remain (at least for now), with benchmarks having been moved to librapidsmpf-tests package to avoid introducing dependencies to libcudf_streaming in the library package.

…cudf repo)

…df repo)

Restore benchmarks, examples, non-cudf tests, sanitizer suppressions, and shuffler header that were deleted outside the intended cudf-specific removal scope. Only streaming/cudf and integrations/cudf directories were supposed to be removed.

After moving cudf-specific streaming code to the cudf_streaming package, update the rapidsmpf benchmarks, tests, and examples that use this functionality to include the new <cudf_streaming/...> headers instead of the removed <rapidsmpf/streaming/cudf/...> headers.

… tests

Remove get_cudf.cmake and cudf::cudf link from core CMakeLists.txt. Replace the single cudf::detail::memcpy_async call in cuda_memcpy_async.hpp with a direct cudaMemcpyAsync call — the cudf wrapper is identical behavior.

Replace all cudf:: stream/memory-resource getters with RMM equivalents: - cudf::get_default_stream() -> rmm::cuda_stream_view{} - cudf::get_current_device_resource_ref() -> rmm::mr::get_current_device_resource_ref() - cudf::test::BaseFixtureWithParam -> ::testing::TestWithParam directly Remove dead cudf_test includes and cudf::cudftestutil link from CMakeLists.

Delete bench_pack.cpp and random_data.{hpp,cu} which were already moved to the cudf repo. Replace cudf::get_default_stream() and cudf::get_current_device_resource_ref() in bench_comm.cpp with RMM equivalents. Clean CMakeLists.txt to remove the deleted targets.

Remove from_cudf_packed_columns classmethod from PackedData (moved to cudf_streaming in cudf repo). Delete testing.py (unused) and streaming_benchmark.py (moved to cudf repo). Remove libcudf.load_library() call from librapidsmpf/load.py.

pentschev

Thanks @vyasr . I did an initial pass and left a few comments. In particular I'm a bit concerned about removing certain benchmarks/examples and more importantly tests that we had for "core" components both in C++ and Python. I think some of them were not necessarily intended to be moved to cudf_streaming, but maybe I'm missing some context. For tests in particular, I think we should still keep core components being tests against cuDF, as they provide important coverage that we would lose if they're not handled here. Would be also good to hear @madsbk 's opinion (and Lawrence too, but I'm not tagging him since he's on PTO).

pentschev · 2026-06-05T17:40:03Z

    --exclude libkvikio.so \
    --exclude libnvcomp.so.5 \


I think kvikio and nvcomp were here for cuDF, so they should probably be removed too.

pentschev · 2026-06-05T17:40:57Z

    CPM_ARGS
    GIT_REPOSITORY https://github.com/NVIDIA/cuCascade.git
-    GIT_TAG main
+    GIT_TAG 6a8cf0f7c545c601cec4b41a0da22ad6a17eeb7c


Why is this pinning to a specific commit and not main?

pentschev · 2026-06-05T17:42:36Z

-  COMPONENT benchmarking
-  DESTINATION bin/benchmarks/librapidsmpf
-  EXCLUDE_FROM_ALL
-)


Why are we removing benchmarks? I think they're not necessarily part of the move to cudf_streaming, in particular those that are not part of the streaming submodule, are they?

I was thinking the same. Cant benchmarks depend on cudf_streaming?

If I'm reading this right, it's still there, but moved down below into a conditional block if(BUILD_CUDF_TESTS).

pentschev · 2026-06-05T17:50:36Z

I'm not sure we want to remove Shuffler tests entirely. We probably want to find a middle-ground here, since removing this file completely will leave Shuffler pretty much untested. Maybe for benchmarks tests we should still link to cuDF since they're not packaged and keep them mostly "as is"?

Remove all depends_on_libcudf and depends_on_pylibcudf entries from dependencies.yaml, conda recipes, pyproject.toml, and conda environments. Remove libcudf.so auditwheel excludes from wheel build scripts. Remove cudf/libcudf/pylibcudf from version-bump script. Remove cudf Doxygen tagfile, intersphinx mapping, and stale API docs for removed integrations module. Remove cudf ASAN/compute-sanitizer suppressions. Remove cudf from clang-format include priority and ruff isort sections. Update docs to reflect cudf is no longer a direct dependency.

…id shell splitting

…atchAsync support

… conda recipe

…udf_streaming CPM

pentschev · 2026-06-06T07:22:13Z

Possibly this and bench_partition.cpp could move to cudf_streaming, they're primarily cuDF tests with no collectives, I don't think it has to be in this PR though. WDYT @nirandaperera ?

pentschev · 2026-06-06T09:46:20Z

Seems like the change in permissions is wanted since execution is removed from this file that contains a shebang.

Reverted this change.

… deps bench_read and other NDSH benchmarks link cudf_streaming::cudf_streaming which transitively pulls in libcudf.so. Since benchmarks are installed into the librapidsmpf package (--component=benchmarking), the package needs these runtime dependencies declared to pass the overlinking check.

…orts Tests and examples import cudf partition/streaming APIs that moved from rapidsmpf.integrations.cudf and rapidsmpf.streaming.cudf to cudf_streaming.integrations and cudf_streaming.streaming respectively. - Add depends_on_pylibcudf and depends_on_cudf_streaming to test_python and py_test_rapidsmpf file keys - Remap all Python imports to cudf_streaming namespace - Regenerate pyproject.toml via rapids-dependency-file-generator

… cudf-streaming conda-only dep

…brapidsmpf-tests

bdice · 2026-06-08T03:15:58Z

+# This warning needs to be suppressed because some parts of rapidsmpf instantiate templated CCCL
 # functions in contexts where the resulting instantiations would have internal linkage (e.g. in
 # anonymous namespaces). In such contexts, the visibility attribute on the template is ignored, and
 # the compiler issues a warning. This is not a problem and will be fixed in future versions of CCCL.


Follow up: This suppression has been here since @madsbk's "initial commit" 2 years ago. We should re-evaluate if it is still needed and remove it if not.

Added this task to NVIDIA/cudf#22771.

bdice · 2026-06-08T03:21:05Z

          - ${{ stdlib("c") }}
        by_name:
          - cuda-cupti
-          - libcudf


We still need to ignore libcudf (and possibly other dependencies? Please double-check the list below). This is now being included from cache-host, because libcudf has a run-export on itself. https://github.com/rapidsai/rapidsmpf/actions/runs/27112901660/job/80014151119?pr=1083#step:11:1401

│ │ Finalized run dependencies (librapidsmpf-26.08.00a74-cuda13_260608_c925c61f): │ │ ╭──────────────────┬──────────────────────────────────────────────────────╮ │ │ │ Name ┆ Spec │ │ │ ╞══════════════════╪══════════════════════════════════════════════════════╡ │ │ │ Run dependencies ┆ │ │ │ │ cuda-cupti ┆ * │ │ │ │ cuda-version ┆ >=13,<14.0a0 (PC) │ │ │ │ libcudf ┆ >=26.8.0a615,<26.9.0a0 (RE of [cache-host: libcudf]) │ │ │ │ libnuma ┆ >=2.0.18,<3.0a0 (RE of [host: libnuma]) │ │ │ │ ┆ >=2.0.18,<3.0a0 (RE of [cache-host: libnuma]) │ │ │ │ libpmix ┆ >=5.0,<6.0 │ │ │ │ ┆ >=5.0.8,<6.0a0 (RE of [host: libpmix-devel]) │ │ │ │ ┆ >=5.0.8,<6.0a0 (RE of [cache-host: libpmix-devel]) │ │ │ │ librmm ┆ 26.8.* │ │ │ │ libucxx ┆ 0.51.* │ │ │ │ ┆ >=0.51.0a25,<0.52.0a0 (RE of [host: libucxx]) │ │ │ │ ┆ >=0.51.0a25,<0.52.0a0 (RE of [cache-host: libucxx]) │ │ │ │ openmpi ┆ >=5.0 │ │ │ ╰──────────────────┴──────────────────────────────────────────────────────╯

Only the librapidsmpf-tests package should have libcudf / libcudf-streaming dependencies. https://github.com/rapidsai/rapidsmpf/actions/runs/27112901660/job/80014151119?pr=1083#step:11:1980

│ │ Finalized run dependencies (librapidsmpf-tests-26.08.00a74-cuda13_260608_c925c61f): │ │ ╭───────────────────┬──────────────────────────────────────────────────────╮ │ │ │ Name ┆ Spec │ │ │ ╞═══════════════════╪══════════════════════════════════════════════════════╡ │ │ │ Run dependencies ┆ │ │ │ │ cuda-cupti ┆ * │ │ │ │ cuda-version ┆ >=13,<14.0a0 (PC) │ │ │ │ libcudf ┆ 26.8.* │ │ │ │ ┆ >=26.8.0a615,<26.9.0a0 (RE of [host: libcudf]) │ │ │ │ ┆ >=26.8.0a615,<26.9.0a0 (RE of [cache-host: libcudf]) │ │ │ │ libcudf-streaming ┆ 26.8.* │ │ │ │ libnuma ┆ >=2.0.18,<3.0a0 (RE of [host: libnuma]) │ │ │ │ ┆ >=2.0.18,<3.0a0 (RE of [cache-host: libnuma]) │ │ │ │ libpmix ┆ >=5.0.8,<6.0a0 (RE of [cache-host: libpmix-devel]) │ │ │ │ librmm ┆ 26.8.* │ │ │ │ libucxx ┆ 0.51.* │ │ │ │ ┆ >=0.51.0a25,<0.52.0a0 (RE of [cache-host: libucxx]) │ │ │ │ openmpi ┆ >=5.0 │ │ │ ╰───────────────────┴──────────────────────────────────────────────────────╯

Good catch.

bdice · 2026-06-08T03:23:15Z

      script:
        content: |
          cmake --install cpp/build --component testing
+          cmake --install cpp/build --component=benchmarking


This has an = but the line above doesn't. Let's be consistent.

Fixed — using --component=testing now for consistency.

bdice · 2026-06-08T03:24:29Z

+      - libcudf-streaming =${{ minor_version }}
+      - libcudf =${{ minor_version }}


We should alphabetize these (libcudf before libcudf-streaming). Please apply throughout.

Fixed — libcudf before libcudf-streaming throughout the recipe.

bdice · 2026-06-08T03:25:04Z

            }
-            if (msg.holds<rapidsmpf::streaming::TableChunk>()) {
-                auto chunk = co_await msg.release<rapidsmpf::streaming::TableChunk>()
+            if (msg.holds<cudf_streaming::streaming::TableChunk>()) {


It's a little odd to have "streaming" twice in the name, maybe this can be refactored later.

Renaming the namespaces is tracked in NVIDIA/cudf#22771.

bdice · 2026-06-08T03:44:32Z

-  COMPONENT benchmarking
-  DESTINATION bin/benchmarks/librapidsmpf
-  EXCLUDE_FROM_ALL
-)


If I'm reading this right, it's still there, but moved down below into a conditional block if(BUILD_CUDF_TESTS).

bdice · 2026-06-08T03:48:44Z

  - Regex: '^<(thrust|cub|cuda)/'
    Priority: 40 # CCCL includes
-  - Regex: '^<(cudf.*|rmm|cugraph|cuml|raft|kvikio|cucascade)'
+  - Regex: '^<(rmm|cugraph|cuml|raft|kvikio|cucascade)'


This change seems unexpected to me. Why remove cudf here? We still have a few optional includes, in the tests/benchmarks. We try to keep this file mostly-consistent across RAPIDS, which is why it mentions other libraries too.

You're right, tests and benchmarks still #include <cudf/...>. Added cudf back to the RAPIDS includes regex.

bdice · 2026-06-08T03:49:56Z

 message(STATUS "  RAPIDSMPF_VERBOSE_INFO  : ${RAPIDSMPF_VERBOSE_INFO}")
+message(STATUS "  BUILD_CUDF_TESTS          : ${BUILD_CUDF_TESTS}")


Spacing.

Suggested change

message(STATUS " RAPIDSMPF_VERBOSE_INFO : ${RAPIDSMPF_VERBOSE_INFO}")

message(STATUS " BUILD_CUDF_TESTS : ${BUILD_CUDF_TESTS}")

message(STATUS " RAPIDSMPF_VERBOSE_INFO : ${RAPIDSMPF_VERBOSE_INFO}")

message(STATUS " BUILD_CUDF_TESTS : ${BUILD_CUDF_TESTS}")

Fixed, thanks.

bdice · 2026-06-08T03:52:41Z

+              - *pylibcudf_unsuffixed
+  depends_on_cudf_streaming:
+    common:
+      - output_types: conda


Are there wheels for cudf-streaming yet? I thought I approved some code for that but maybe it wasn't merged at the time this was written. Please check.

Not yet. We should have had them after NVIDIA/cudf#22760, but there was an oversight in that PR's changes to build.yaml — it didn't properly specify that Python artifacts of cudf_streaming use the stable ABI, which causes failures like this one. The circularity of the rapidsmpf↔cudf dependency tree (partly alleviated by this branch making cudf a test-only dep of rapidsmpf) meant we couldn't merge a fix there until this rapidsmpf branch lands. Once this merges we'll get that fixed and enable the wheels.

…IDS regex

bdice

Thank you. Approving packaging and CI and CMake.

pentschev · 2026-06-08T09:51:50Z

It seems Niranda's concerns (like mine) have been addressed, so I'm going ahead and merging it to speedup unblocking.

pentschev · 2026-06-08T09:51:52Z

/merge

Concerns should have been addressed. Please open an issue if something is still not addressed.

This PR removes the `-diag-suppress=1407` NVCC flag from `cmake/Modules/ConfigureCUDA.cmake`. ## Background The suppression was added in the initial commit ~2 years ago to silence NVCC warning 1407, which fires when CCCL templates are instantiated in contexts with internal linkage (e.g. anonymous namespaces). The comment noted: *"This is not a problem and will be fixed in future versions of CCCL."* The item was tracked as a follow-up in [NVIDIA/cudf#22771](NVIDIA/cudf#22771), referenced from [PR #1083 discussion](#1083 (comment)). ## Why it's safe to remove 1. **The library has zero `.cu` files** — all source under `cpp/src/` is `.cpp`, so the NVCC-specific flag never applied to library compilation. 2. **The only `.cu` file is `cpp/tests/test_allreduce.cu`**, which uses CCCL templates (`cuda::std::plus`, `cuda::minimum`, etc.) inside an anonymous namespace — exactly the pattern that used to trigger the warning. 3. **Verified empirically**: compiling `test_allreduce.cu` with `-Werror=all-warnings` and **without** `-diag-suppress=1407` using CUDA 12.9 (nvcc V12.9.86) produces **zero** warning 1407 instances. CCCL has fixed the underlying visibility attribute issue. ## Addresses - Partial fix for [NVIDIA/cudf#22771](NVIDIA/cudf#22771) (the rapidsmpf follow-up item) Authors: - Vyas Ramasubramani (https://github.com/vyasr) Approvers: - Mads R. B. Kristensen (https://github.com/madsbk) URL: #1100

vyasr self-assigned this Jun 5, 2026

vyasr requested review from a team as code owners June 5, 2026 17:11

vyasr added breaking Introduces a breaking change improvement Improves an existing functionality labels Jun 5, 2026

vyasr requested a review from msarahan June 5, 2026 17:11

vyasr added 5 commits June 5, 2026 17:12

refactor: remove cudf-specific streaming/integrations code (moved to …

698149d

…cudf repo)

refactor: remove cudf-specific Python bindings and tests (moved to cu…

f4908d9

…df repo)

refactor: clean up packaging after cudf streaming code removal

8ed93f8

vyasr force-pushed the chore/remove_cudf_components branch from b9e7f2b to 8d136d0 Compare June 5, 2026 17:13

rapidsai deleted a comment from copy-pr-bot Bot Jun 5, 2026

vyasr added 10 commits June 5, 2026 17:17

fix: correct include path in shuffler.hpp

0b7e80a

refactor: remove cudf-specific benchmarks (moved to cudf repo)

24539a3

refactor: remove cudf-specific C++ tests, add packed data tests

dabc851

refactor: remove cudf-specific C++ examples (moved to cudf repo)

dc41604

refactor: remove cudf-specific Python tests/examples, add packed data…

11eb473

… tests

Use topology-only build of cucascade

4dc1c40

vyasr force-pushed the chore/remove_cudf_components branch from 8d136d0 to 21a841d Compare June 5, 2026 17:44

pentschev reviewed Jun 5, 2026

View reviewed changes

vyasr added 2 commits June 5, 2026 23:05

fix(packaging): use spaces instead of semicolons in cmake-args to avo…

d120f5d

…id shell splitting

feat(memory): add cuda_memcpy_batch_async with CUDA 13.0+ cudaMemcpyB…

7e5d658

…atchAsync support

vyasr force-pushed the chore/remove_cudf_components branch from cca7c83 to 7e5d658 Compare June 5, 2026 23:06

vyasr mentioned this pull request Jun 5, 2026

Use cuCascade topology-only build mode #1080

Closed

vyasr added 4 commits June 5, 2026 23:14

fix(packaging): use escaped quotes with space-separated cmake args in…

bd34e91

… conda recipe

fix(packaging): match cudf recipe pattern for multiple cmake-args values

c6d9a04

fix(cmake): guard rapidsmpf::rapidsmpf alias to avoid conflict with c…

46d3af6

…udf_streaming CPM

docs(cmake): add comment explaining rapidsmpf::rapidsmpf alias guard

42562a3

pentschev reviewed Jun 6, 2026

View reviewed changes

vyasr added 8 commits June 6, 2026 14:12

fix(cudf): update restored tests and benchmarks for cudf_streaming APIs

a397198

fix(packaging): remove libcudf-streaming from librapidsmpf host, make…

f509d62

… cudf-streaming conda-only dep

fix(tests): skip cudf_streaming tests gracefully when wheel unavailable

85e4ae9

fix(packaging): move benchmarks and cudf deps from librapidsmpf to li…

21ee1b6

…brapidsmpf-tests

fix(tests): restore rapidsmpf.testing module needed by test suite

8879441

Revert permissions

c925c61

bdice reviewed Jun 8, 2026

View reviewed changes

fix(review): address bdice review feedback

1b4ee3c

vyasr mentioned this pull request Jun 8, 2026

Follow-ups for rapidsmpf to cudf migration NVIDIA/cudf#22771

Open

10 tasks

style: apply clang-format include reordering after adding cudf to RAP…

e644a81

…IDS regex

bdice approved these changes Jun 8, 2026

View reviewed changes

pentschev approved these changes Jun 8, 2026

View reviewed changes

pentschev requested a review from nirandaperera June 8, 2026 09:51

rapids-bot Bot merged commit 3cba8e9 into rapidsai:main Jun 8, 2026
135 of 167 checks passed

vyasr mentioned this pull request Jun 13, 2026

build: remove stale NVCC -diag-suppress=1407 #1100

Merged

		- libcudf-streaming =${{ minor_version }}
		- libcudf =${{ minor_version }}

		message(STATUS " RAPIDSMPF_VERBOSE_INFO : ${RAPIDSMPF_VERBOSE_INFO}")
		message(STATUS " BUILD_CUDF_TESTS : ${BUILD_CUDF_TESTS}")

Uh oh!

Conversation

vyasr commented Jun 5, 2026 • edited by pentschev Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pentschev left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

bdice left a comment

Choose a reason for hiding this comment

Uh oh!

pentschev commented Jun 8, 2026

Uh oh!

pentschev commented Jun 8, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

vyasr commented Jun 5, 2026 •

edited by pentschev

Loading