[CI] SG_SVC_TEST failed

[aamijar]

May 6th - 7th Nightly CI Failure
https://github.com/rapidsai/cuml/actions/runs/25483389976/job/74772823505

This is likely related to the following CCCL bug:
NVIDIA/cccl#8838
NVIDIA/cccl#8839 - fix merged May 6th.

NVIDIA/cuopt#1184 - cuOpt stopgap
NVIDIA/cuvs#2062 - cuVS stopgap

cuML CI is not picking up the new fix yet.

This may be fixed once rapidsai/rapids-cmake#1018 is merged.

2026-05-07T08:01:38.6936445Z [  FAILED  ] 3 tests, listed below:
2026-05-07T08:01:38.6939624Z [  FAILED  ] SmoSolverTest/0.SvcTest, where TypeParam = float
2026-05-07T08:01:38.6941200Z [  FAILED  ] SmoSolverTest/0.DenseBatching, where TypeParam = float
2026-05-07T08:01:38.6942703Z [  FAILED  ] SmoSolverTest/0.SparseBatching, where TypeParam = float

2026-05-07T08:01:38.6645936Z [ RUN      ] SmoSolverTest/0.SvcTest
2026-05-07T08:01:38.6648409Z unknown file: Failure
2026-05-07T08:01:38.6650775Z C++ exception with description "exception occurred! file=/tmp/conda-bld-output/bld/rattler-build_libcuml/work/cpp/tests/../src/svm/results.cuh line=410: Incorrect training: cannot calculate the constant in the decision function

2026-05-07T08:01:38.6732211Z [ RUN      ] SmoSolverTest/0.DenseBatching
2026-05-07T08:01:38.6733376Z unknown file: Failure
2026-05-07T08:01:38.6735862Z C++ exception with description "exception occurred! file=/tmp/conda-bld-output/bld/rattler-build_libcuml/work/cpp/tests/../src/svm/results.cuh line=410: Incorrect training: cannot calculate the constant in the decision function

2026-05-07T08:01:38.6796865Z [ RUN      ] SmoSolverTest/0.SparseBatching
2026-05-07T08:01:38.6800047Z unknown file: Failure
2026-05-07T08:01:38.6802251Z C++ exception with description "exception occurred! file=/tmp/conda-bld-output/bld/rattler-build_libcuml/work/cpp/src/svm/svc_impl.cuh line=152: Only binary classification is implemented at the moment

Source code in src/svm/results.cuh that uses cub

cub::DeviceSelect::Flagged(cub_storage.data(),
                               cub_bytes,
                               f,
                               flag.data(),
                               val_selected.data(),
                               d_num_selected.data(),
                               n_train,
                               stream);
    int n_selected;
    raft::update_host(&n_selected, d_num_selected.data(), 1, stream);
    handle.sync_stream(stream);
    math_t res = 0;
    ASSERT(n_selected > 0,
           "Incorrect training: cannot calculate the constant in the decision "
           "function");

[csadorf]

Additional occurrence of the same SG_SVC_TEST failure observed on PR #8069:

• PR job: https://github.com/rapidsai/cuml/actions/runs/25561326428/job/75035402871?pr=8069
• Job name: conda-cpp-tests / 12.2.2, 3.11, arm64, ubuntu22.04, a100, latest-driver, latest-deps
• SHA: c0b6ba6f8518643bfab61ceb086265d13d6a9cb7

The same failure is also present on main:

• Main job: https://github.com/rapidsai/cuml/actions/runs/25483389976/job/74772823506
• Job name: conda-cpp-tests / 13.1.1, 3.13, arm64, rockylinux8, a100, latest-driver, latest-deps
• SHA: 83cd3f8e1cd9a17ac69f5d520b78c01fe7c262aa

Both jobs fail in SG_SVC_TEST with the same three failing tests:

SmoSolverTest/0.SvcTest, where TypeParam = float
SmoSolverTest/0.DenseBatching, where TypeParam = float
SmoSolverTest/0.SparseBatching, where TypeParam = float

The observed error messages match this issue:

results.cuh line=410:
Incorrect training: cannot calculate the constant in the decision function

svc_impl.cuh line=152:
Only binary classification is implemented at the moment

In the PR #8069 job, the FIL buffer tests passed before SG_SVC_TEST failed:

SG_HOST_BUFFER_TEST ... Passed
SG_DEVICE_BUFFER_TEST ... Passed

This appears to be the same pre-existing SG_SVC_TEST failure tracked here, not a regression from PR #8069.

[jinsolp]

Also observed in this PR run https://github.com/rapidsai/cuml/actions/runs/25518669067/job/74897430680

[divyegala]

Any chance it could be related to #6569 and how that PR triaged it down to CCCL updates?