Describe the bug
Building libcuopt from source against rmm 26.10 fails to compile. rmm 26.10 places device_uvector inside an ABI inline namespace (rmm::RMM_ABI_NAMESPACE, see rmm/detail/export.hpp), while cuOpt hand-rolls a forward declaration of it in the plain rmm namespace in cpp/src/pdlp/utilities/problem_checking.cuh:
namespace rmm {
template <typename T>
class device_uvector;
} // namespace rmmThis injects a second, distinct rmm::device_uvector, so every use becomes ambiguous.
Steps/Code to reproduce
Configure and build libcuopt from source with rmm 26.10 on the include path (CUDA 13.3 / GCC 14 / sm_120a here, but the ambiguity is compiler/arch independent).
Error
cpp/src/pdlp/utilities/problem_checking.cuh(36): error: "rmm::device_uvector" is ambiguous
cpp/src/utilities/copy_helpers.hpp(82): error: "rmm::device_uvector" is ambiguous
...
The failures then cascade into follow-on namespace "cuopt::raft" has no member "add_op" and thrust "this pragma must immediately precede a declaration" errors in the same translation units; all of them trace back to this one forward declaration (fixing it alone clears them).
Root cause
problem_checking.cuh is the only place in the tree that manually forward-declares an rmm type. cuOpt already opts out of thrust's and cub's ABI namespaces (-DTHRUST_DISABLE_ABI_NAMESPACE / -DCUB_DISABLE_NAMESPACE_MAGIC); rmm's ABI inline namespace is newer and has no such opt-out, so the plain forward declaration no longer matches rmm's real declaration.
Fix
Replace the hand-rolled forward declaration with #include <rmm/device_uvector.hpp> (rmm's own, correctly ABI-namespaced declaration). PR to follow.
Environment
- cuOpt:
main
- rmm / raft: 26.10
- CCCL: 3.4.0, CUDA 13.3 toolkit, GCC 14
- GPU: RTX 5090 (
sm_120a)
Describe the bug
Building
libcuoptfrom source against rmm 26.10 fails to compile. rmm 26.10 placesdevice_uvectorinside an ABI inline namespace (rmm::RMM_ABI_NAMESPACE, seermm/detail/export.hpp), while cuOpt hand-rolls a forward declaration of it in the plainrmmnamespace incpp/src/pdlp/utilities/problem_checking.cuh:This injects a second, distinct
rmm::device_uvector, so every use becomes ambiguous.Steps/Code to reproduce
Configure and build
libcuoptfrom source with rmm 26.10 on the include path (CUDA 13.3 / GCC 14 /sm_120ahere, but the ambiguity is compiler/arch independent).Error
The failures then cascade into follow-on
namespace "cuopt::raft" has no member "add_op"and thrust"this pragma must immediately precede a declaration"errors in the same translation units; all of them trace back to this one forward declaration (fixing it alone clears them).Root cause
problem_checking.cuhis the only place in the tree that manually forward-declares an rmm type. cuOpt already opts out of thrust's and cub's ABI namespaces (-DTHRUST_DISABLE_ABI_NAMESPACE/-DCUB_DISABLE_NAMESPACE_MAGIC); rmm's ABI inline namespace is newer and has no such opt-out, so the plain forward declaration no longer matches rmm's real declaration.Fix
Replace the hand-rolled forward declaration with
#include <rmm/device_uvector.hpp>(rmm's own, correctly ABI-namespaced declaration). PR to follow.Environment
mainsm_120a)