fix(rocm): install libcuopt.so and libmps_parser.so into the wheels

[awelling2801]

Problem

Installing the published ROCm wheels (rocopt-mps-parser, librocopt, rocopt, rocopt-server, rocopt-sh-client) from pypi.amd.com and running import cuopt fails on a clean environment:

OSError: libmps_parser.so: cannot open shared object file: No such file or directory

Even if that's patched around, the next failure is the same thing for libcuopt.so. Neither library is actually present anywhere in the published wheels.

Root cause

Two separate missing install() rules, both ROCm-specific:

1. python/libcuopt/CMakeLists.txt — the USE_ROCM branch returns early with zero install rules:

if(USE_ROCM)
    ...
    return()  # nothing installed, ever
endif()

libcuopt.so is already compiled by the C++ build step (cpp/build/libcuopt.so), but this function never copies it into the wheel. load.py's _load_wheel_installation() looks for it at <package>/lib64/libcuopt.so — exactly matching the layout of NVIDIA's own CUDA wheel (libcuopt-cu12, which bundles libcuopt/lib64/libcuopt.so) — but that path is never populated on the ROCm side. This is why the project's own Docker image (dockerfile.rocm) works around it by hardcoding cpp/build/ onto LD_LIBRARY_PATH instead of relying on the wheel being self-contained.

2. python/cuopt/cuopt/linear_programming/data_model/CMakeLists.txt — data_model_wrapper's compiled extension links against the cuopt::mps_parser IMPORTED target but never installs it:

set(linked_libraries cuopt::mps_parser)
rapids_cython_create_modules(... LINKED_LIBRARIES "${linked_libraries}" ...)
# no install_aliased_imported_targets() call

WheelHelpers.cmake already defines install_aliased_imported_targets() for exactly this purpose ("Making libraries available inside wheels by installing the associated targets") — it's just never called anywhere in the repo, on either the CUDA or ROCm path.

Fix

• Add the missing install(FILES ...) call for libcuopt.so in python/libcuopt/CMakeLists.txt, with a FATAL_ERROR guard if the prebuilt .so doesn't exist yet (fail loudly at build time instead of silently shipping an empty wheel).
• Add the missing install_aliased_imported_targets(TARGETS cuopt::mps_parser DESTINATION cuopt/linear_programming/data_model) call so libmps_parser.so lands next to data_model_wrapper's compiled module, matching its INSTALL_RPATH=$ORIGIN.

Verification

Confirmed against the actual published wheels at pypi.amd.com/rocm-7.2.3/packages/rocopt/:

• librocopt-1.0.0 contains no compiled binaries at all (full wheel content listing, ~10KB total).
• rocopt-1.0.0's data_model_wrapper*.so has RPATH=$ORIGIN per readelf -d, with no libmps_parser.so anywhere in the wheel.
• Manually placing libmps_parser.so at that exact path (as a local shim wheel layered on top) was confirmed to resolve the import error, isolating this as the precise root cause.

I wasn't able to do a full ROCm build in this environment to produce a rebuilt wheel end-to-end, so this PR is scoped to the CMake install-rule fix itself; CI should pick it up on rebuild.

Related

Also checked python/cuopt/cuopt/linear_programming/internals/CMakeLists.txt, which has the same cuopt::cuopt linking pattern without an explicit install. This one appears to be fine as-is: cuopt/__init__.py already does import libcuopt; libcuopt.load_library() before importing internals/linear_programming, which preloads libcuopt.so into the process via ctypes.CDLL ahead of time (the same pattern already used for librmm/libraft/rapids_logger). Once libcuopt.so actually ships (fix #1 above), that preload should let internals's DT_NEEDED=libcuopt.so resolve against the already-loaded instance without needing its own bundled copy.

Made with Cursor