Version

26.06 (main branch)

Which installation method(s) does this occur on?

Pip

Describe the bug.

The test_ktruss_subgraph_Graph test fails on Blackwell (RTX PRO 6000, sm_120) when built with nvcc 13.0 + CCCL 3.4.0. The bug was caught on CI through the wheel test configuration which uses nvcc 13.0, but the issue is not wheel-specific and affects any build using that nvcc version on Blackwell. The failure manifests as CUGRAPH_UNKNOWN_ERROR: edge_list contains edges that do not exist in the input graph during the k-truss peeling loop. The root cause is thrust::reduce_by_key on zip_iterator<tuple<int32_t, int32_t>> silently corrupting its output.

The bug is in nvcc 13.0 and is fixed in nvcc 13.2. Both use the exact same CCCL 3.4.0 headers; the only delta is the nvcc binary.

A workaround has been applied in k_truss_impl.cuh [MENTION PR NUMBER] replacing thrust::reduce_by_key with thrust::unique_copy + thrust::lower_bound / thrust::upper_bound.

The original reduce_by_key call should be restored once nvcc 13.0 is no longer supported or the nvcc fix is confirmed backported.

Minimum reproducible example

Run the python `test_ktruss_subgraph_Graph` with the above configurations.

Relevant log output

RuntimeError: non-success value returned from cugraph_k_truss_subgraph: CUGRAPH_UNKNOWN_ERROR
cuGraph failure at file=cpp/include/cugraph/prims/transform_e.cuh line=589:
Invalid input arguments: edge_list contains edges that do not exist in the input graph.

Environment details

GPU: NVIDIA RTX PRO 6000 Blackwell Server Edition (sm_120)
Driver: 580.95.05
CUDA Toolkit: 13.0.88 (fails) / 13.2.51 (passes)
CCCL: 3.4.0 (same for both)
OS: Ubuntu 24.04 (container)
CI config: wheel-tests-cugraph / 13.1.1, 3.14, amd64, ubuntu24.04, rtxpro6000

Other/Misc.

No response

Code of Conduct

• I agree to follow cuGraph's Code of Conduct
• I have searched the open bugs and have found no duplicates for this bug report