Build and test with CUDA 12.8.0

[bdice]

This PR uses CUDA 12.8.0 to build and test.

xref: rapidsai/build-planning#139

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[jameslamb]

It looks like there are more code changes required here.

Error: calling a __host__ function from a __host__ __device__ function is not allowed
          detected during:
            instantiation of "cuda::std::__4::__optional_destruct_base<_Tp, true>::__optional_destruct_base(cuda::std::__4::in_place_t, _Args &&...) [with _Tp=cugraph::detail::edge_partition_edge_property_device_view_t<int32_t, const uint32_t *, __nv_bool>, _Args=<cugraph::edge_property_view_t<int32_t, const uint32_t *, __nv_bool> &, int>]" at line 705 of $SRC_DIR/cpp/build/_deps/cccl-src/libcudacxx/lib/cmake/libcudacxx/../../../include/cuda/std/detail/libcxx/include/optional
            implicit generation of "cuda::std::__4::__optional_storage_base<cuda::std::__4::decay<cugraph::detail::edge_partition_edge_property_device_view_t<cugraph::graph_view_t<int32_t, int32_t, false, false, void>::edge_type, const uint32_t *, __nv_bool>>::type>::__optional_storage_base(cuda::std::__4::in_place_t, _Args &&...) [with _Args=<cugraph::edge_property_view_t<int32_t, const uint32_t *, __nv_bool> &, int>]" at line 705 of $SRC_DIR/cpp/build/_deps/cccl-src/libcudacxx/lib/cmake/libcudacxx/../../../include/cuda/std/detail/libcxx/include/optional
            instantiation of class "cuda::std::__4::__optional_storage_base<_Tp> [with _Tp=cugraph::detail::edge_partition_edge_property_device_view_t<int32_t, const uint32_t *, __nv_bool>]" at line 705 of $SRC_DIR/cpp/build/_deps/cccl-src/libcudacxx/lib/cmake/libcudacxx/../../../include/cuda/std/detail/libcxx/include/optional
            implicit generation of "cuda::std::__4::__optional_copy_base<cuda::std::__4::decay<cugraph::detail::edge_partition_edge_property_device_view_t<cugraph::graph_view_t<int32_t, int32_t, false, false, void>::edge_type, const uint32_t *, __nv_bool>>::type, cuda::std::__4::__smf_availability::__trivial>::__optional_copy_base(cuda::std::__4::in_place_t, _Args &&...) [with _Args=<cugraph::edge_property_view_t<int32_t, const uint32_t *, __nv_bool> &, int>]" at line 705 of $SRC_DIR/cpp/build/_deps/cccl-src/libcudacxx/lib/cmake/libcudacxx/../../../include/cuda/std/detail/libcxx/include/optional
            instantiation of class "cuda::std::__4::__optional_copy_base<_Tp, <unnamed>> [with _Tp=cugraph::detail::edge_partition_edge_property_device_view_t<int32_t, const uint32_t *, __nv_bool>, <unnamed>=cuda::std::__4::__smf_availability::__trivial]" at line 705 of $SRC_DIR/cpp/build/_deps/cccl-src/libcudacxx/lib/cmake/libcudacxx/../../../include/cuda/std/detail/libcxx/include/optional
            [ 7 instantiation contexts not shown ]

(conda build link (CUDA 12.8))

I see similar stuff in CUDA 12.8 wheel builds.

[bdice]

I'm pretty sure builds were expected to work here. This could be a new-ish conflict from #4891 interacting poorly with CUDA 12.8. @miscco Do you have any thoughts on this?

[miscco]

Oh screw it, it seems we are missing the execution check suppression on the bae class constructors

[jakirkham]

Think this is fixed with Rob's PR: rapidsai/rapids-cmake#756

This backports Michael's CCCL PR: NVIDIA/cccl#3578

[bdice]

Wheel builds are only like ~10-13% done after 4 hours. This seems unusual, even with the new compiler and new CCCL version causing zero sccache hits.

[vyasr]

/merge

[jakirkham]

One CI job failed due to some SSL error

FAILED tests/utils/test_dataset.py::test_benchmarking_datasets[dataset3] - ssl.SSLError: [SYS] unknown error (_ssl.c:2578)

Restarting CI

[jakirkham]

Huzzah! 🥳

Thanks all 🙏