build: add the CUDA-free cuopt_client library

[ramakrishnap-nv]

4 of 4 toward a CUDA-free client library. Stacked on #1803. This is the payoff PR — almost entirely CMake, since #1801–#1803 did the code work.

Why

Calling a remote cuopt_grpc_server currently requires the full GPU stack. pip install cuopt pulls cudf, cupy-cuda13x[ctk], rmm, pylibraft, numba-cuda, scipy, pandas and libcuopt onto a machine that only serializes protobuf over a socket.

The motivating consumer is an MCP server: it imports only Client, TlsConfig, DataModel, Read, SolverSettings — no Solve — yet installs the entire CUDA stack today.

What

cuopt_client holds the host-side problem representation (parsers, data_model_view, mps_data_model, writers), the gRPC wire protocol (generated protos + mappers) and the gRPC clients for both LP/MIP and routing. libcuopt and cuopt_grpc_server both link it — one mapper implementation, not a client-side fork.

Both gRPC arms qualify: the routing mappers from #1597 reference no raft/rmm/thrust either. solve_remote.cpp stays in libcuopt — it's the local-vs-remote dispatcher and calls into the GPU solver.

Two build details worth reviewing

Object + shared library pair, mirroring cuopt_objs/cuopt. cuopt_static embeds the objects directly rather than linking the shared library, because internal test binaries reach parser internals like mps_phase_registry_t. Linking the shared library instead produces link failures in MPS_FAST_PARSER_TEST, GRPC_CLIENT_TEST and PDLP_MG_TEST.

Default visibility, deliberately unlike cuopt_objs. cuopt_objs can hide everything not marked CUOPT_EXPORT because libcuopt has a curated public API. cuopt_client is different — it was carved out of the internals, so libcuopt depends on ~214 of its symbols (essentially the whole cpu_optimization_problem_t / data_model_view_t / mps_data_model_t / grpc_client_t surface). With hidden visibility, libcuopt.so fails to load:

undefined symbol: cuopt::mathematical_optimization::grpc_client_t::solve_mip<int,double>

Curating that would mean annotating essentially every host-side method. Open to opinions, but default visibility looks like the right trade.

Also: logger.cpp moves into the client sources — both the parsers and the gRPC code include <utilities/logger.hpp>.

Result

libcuopt_client.so  3.1 MB
  NEEDED : libgrpc++, libprotobuf, libabseil, librapids_logger,
           libc, libstdc++, libgomp, libdl
           -> no libcudart, no rmm, no raft, no cudss
  ldd -r over libcuopt.so + libcuopt_client.so : 0 unresolved symbols

Known gaps

• 14 cuopt::routing:: symbols still undefined — host-only accessors of routing::solver_settings_t / assignment_t in routing/solver_settings.cu and assignment.cu. Same pattern as refactor: split host-only members out of CUDA translation units #1801; follow-up.
• No packaging or Python changes yet. Relinking the Cython extensions, splitting the CPU half of solver_wrapper.pyx, and the cuopt-client wheel / libcuopt-client conda output all come next.
• raft/rmm headers are still needed at build time (the CPU headers transitively include them). Costs nothing at runtime because the device getters throw, but a standalone client package would need the headers separated too — that additionally removes 21 throwing stubs from cpu_optimization_problem_t.

Testing

Full build + 126 test binaries, 0 errors. 111/125 pass; the 14 failures are cudaErrorUnknown from a locally wedged nvidia_uvm, identical on unmodified main. The GPU-dependent tests need a clean CI run to confirm.

🤖 Generated with Claude Code

[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.

[ramakrishnap-nv]

/ok to test

[github-actions]

CI Test Summary

⏭️ All 5 test job(s) skipped.