build(cmake): split libcuopt into cuopt_base / cuopt_routing / cuopt_lp component libs

[ramakrishnap-nv]

Splits libcuopt.so into four component SHARED libraries (cuopt_base, cuopt_routing, cuopt_lp, cuopt_grpc) plus a thin umbrella libcuopt.so that links them all — existing -lcuopt consumers are unaffected.

🤖 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 ec81ebc

[ramakrishnap-nv]

@CodeRabbit review

[coderabbitai]

✅ Action performed

Review finished.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[coderabbitai]

📝 Walkthrough

Walkthrough

Changes

cuOpt componentized build and remote execution

Layer / File(s)	Summary
Component source aggregation cpp/src/CMakeLists.txt, cpp/src/.../CMakeLists.txt	CMake now propagates nested source lists and aggregates base, routing, and LP sources for downstream targets.
Component libraries and dependencies cpp/CMakeLists.txt	Shared component targets, NCCL and KaMinPar wiring, umbrella symbol anchors, installation exports, and conditional gRPC dependencies are updated.
Dynamic remote-solver registration cpp/include/.../remote_solve_registry.hpp, cpp/src/grpc/..., cpp/src/pdlp/..., cpp/src/mip_heuristics/solve.cu, cpp/cuopt_cli.cpp	Remote LP and MIP execution uses registered callbacks and loads the gRPC component on demand.
Distribution and examples ci/*, conda/recipes/libcuopt/recipe.yaml, python/libcuopt/CMakeLists.txt, docs/cuopt/...	Packaging expectations, wheel repair, install RPATHs, C example linking, and documentation link checking are updated.

Estimated code review effort: 4 (Complex) | ~60 minutes

Suggested reviewers: chris-maes, iroy30

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly summarizes the primary change: splitting libcuopt into component libraries using CMake.
Description check	✅ Passed	The description accurately explains the component-library split, umbrella library, and compatibility goal.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches 💡 2

📝 Generate docstrings 💡

• Create stacked PR
• Commit on current branch

🛠️ Fix failing CI checks 💡

• Fix failing CI checks

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feat/split-routing-lp-libs

---

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

cpp/CMakeLists.txt (1)

902-918: 🗄️ Data Integrity & Integration | 🟠 Major | ⚡ Quick win

Exported component names should match the documented API

cuopt::base / cuopt::lp are build-tree aliases only. The install/export set will expose the real targets (cuopt::cuopt_base, cuopt::cuopt_lp, etc.) unless those targets set EXPORT_NAME, so a consumer using find_package(cuopt) won’t be able to link against cuopt::base as documented.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@cpp/CMakeLists.txt` around lines 902 - 918, Update the component target
export configuration used by the rapids_export INSTALL and BUILD calls so
installed targets retain the documented names cuopt::base, cuopt::routing,
cuopt::lp, and cuopt::grpc. Set the appropriate EXPORT_NAME values on the
underlying cuopt_component targets, while preserving cuopt::cuopt as the
umbrella target and keeping build-tree aliases consistent.

🧹 Nitpick comments (2)

cpp/CMakeLists.txt (2)

549-556: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

Give the git hash lookup a fallback for non-git source trees.

Without RESULT_VARIABLE/ERROR_QUIET, tarball builds leak git's error to the configure log and bake an empty hash into build_info.hpp.

♻️ Proposed fallback

 execute_process(
         COMMAND git rev-parse --short HEAD
         WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}
         OUTPUT_VARIABLE GIT_COMMIT_HASH
         OUTPUT_STRIP_TRAILING_WHITESPACE
+        RESULT_VARIABLE _git_hash_result
+        ERROR_QUIET
 )
+if(NOT _git_hash_result EQUAL 0 OR GIT_COMMIT_HASH STREQUAL "")
+    set(GIT_COMMIT_HASH "unknown")
+endif()

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@cpp/CMakeLists.txt` around lines 549 - 556, Update the git hash lookup in the
top-level CMake configuration to capture the execute_process result and suppress
stderr for source trees without Git metadata. When the lookup fails, assign a
stable non-empty fallback hash before the existing GIT_COMMIT_HASH message and
build_info.hpp generation; preserve the real short HEAD value for Git checkouts.

---

664-675: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚖️ Poor tradeoff

cuopt_objs duplicates the component include/definition setup.

The object library re-declares papilo/pslp/dejavu includes, CUDSS defines, and architecture defines that cuopt_configure_component (plus cuopt_lp) already establish. Since cuopt_objs backs cuopt_static for the test builds, drift here means tests compile under a different configuration than shipped libraries. Consider factoring the shared include/definition block into a helper both paths call.

Also applies to: 693-694, 720-724

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@cpp/CMakeLists.txt` around lines 664 - 675, Refactor the shared include
directories and compile definitions currently duplicated by cuopt_objs and the
cuopt_configure_component/cuopt_lp setup into a reusable CMake helper. Invoke
that helper for both cuopt_objs and the shipped-library path, including the
papilo/pslp/dejavu includes, CUDSS definitions, and architecture definitions,
while preserving target-specific settings such as POSITION_INDEPENDENT_CODE and
logging definitions.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Outside diff comments:
In `@cpp/CMakeLists.txt`:
- Around line 902-918: Update the component target export configuration used by
the rapids_export INSTALL and BUILD calls so installed targets retain the
documented names cuopt::base, cuopt::routing, cuopt::lp, and cuopt::grpc. Set
the appropriate EXPORT_NAME values on the underlying cuopt_component targets,
while preserving cuopt::cuopt as the umbrella target and keeping build-tree
aliases consistent.

---

Nitpick comments:
In `@cpp/CMakeLists.txt`:
- Around line 549-556: Update the git hash lookup in the top-level CMake
configuration to capture the execute_process result and suppress stderr for
source trees without Git metadata. When the lookup fails, assign a stable
non-empty fallback hash before the existing GIT_COMMIT_HASH message and
build_info.hpp generation; preserve the real short HEAD value for Git checkouts.
- Around line 664-675: Refactor the shared include directories and compile
definitions currently duplicated by cuopt_objs and the
cuopt_configure_component/cuopt_lp setup into a reusable CMake helper. Invoke
that helper for both cuopt_objs and the shipped-library path, including the
papilo/pslp/dejavu includes, CUDSS definitions, and architecture definitions,
while preserving target-specific settings such as POSITION_INDEPENDENT_CODE and
logging definitions.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 2d788dd1-855a-4038-900e-9a3eab8f07be

📥 Commits

Reviewing files that changed from the base of the PR and between 91134a2 and ec81ebc.

📒 Files selected for processing (13)

• conda/recipes/libcuopt/recipe.yaml
• cpp/CMakeLists.txt
• cpp/src/CMakeLists.txt
• cpp/src/barrier/CMakeLists.txt
• cpp/src/branch_and_bound/CMakeLists.txt
• cpp/src/cuts/CMakeLists.txt
• cpp/src/dual_simplex/CMakeLists.txt
• cpp/src/io/CMakeLists.txt
• cpp/src/linear_algebra/CMakeLists.txt
• cpp/src/math_optimization/CMakeLists.txt
• cpp/src/mip_heuristics/CMakeLists.txt
• cpp/src/pdlp/CMakeLists.txt
• cpp/src/routing/CMakeLists.txt

[ramakrishnap-nv]

/ok to test a54e62a

[github-actions]

CI Test Summary

✅ All 31 test job(s) passed.

[ramakrishnap-nv]

/ok to test 705167f

[ramakrishnap-nv]

/ok to test 1bc38f7

[ramakrishnap-nv]

/ok to test 8511bd0

[ramakrishnap-nv]

/ok to test f2b22e7

[ramakrishnap-nv]

/ok to test 90ee992

[ramakrishnap-nv]

/ok to test a83e0f4

[ramakrishnap-nv]

/ok to test 0f3b7e1

[ramakrishnap-nv]

/ok to test b872bfb

[ramakrishnap-nv]

/ok totest 7b32d1b

[ramakrishnap-nv]

/ok to test 7b32d1b

[ramakrishnap-nv]

/ok to test 80c1c34

[ramakrishnap-nv]

@CodeRabbit review

[coderabbitai]

✅ Action performed

Review finished.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[coderabbitai]

Actionable comments posted: 5

🧹 Nitpick comments (2)

cpp/src/mip_heuristics/solve.cu (1)

918-922: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Duplicated lazy-load-and-validate sequence.

This is a line-for-line copy of the LP path in cpp/src/pdlp/solve.cu (Lines 2696-2700), including the discarded dlopen result and the hardcoded unversioned soname. A single helper in the registry (e.g. ensure_remote_solvers_loaded()) would keep the loader name, the dlerror() reporting, and any future synchronization in one place instead of two components.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@cpp/src/mip_heuristics/solve.cu` around lines 918 - 922, Replace the inline
lazy-load and null-check sequence in the MIP remote execution path with a shared
registry helper such as ensure_remote_solvers_loaded(). Move loader-name
handling, dlopen result validation, dlerror reporting, and synchronization into
that helper, then reuse it from both this path and the corresponding LP path
while preserving the existing runtime-error behavior when the remote solver
remains unavailable.

cpp/include/cuopt/mathematical_optimization/remote_solve_registry.hpp (1)

25-30: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

Add Doxygen comments for the new registry surface.

This lives under cpp/include/cuopt/, where new declarations are expected to carry Doxygen-style documentation. Line comments cover the intent, but the callback contract worth stating explicitly is: who owns the returned solution, whether the callbacks may throw across the component boundary, and that only <int, double> is instantiated.

As per path instructions, "New public functions/classes need Doxygen-style documentation" and "Suggest documenting thread-safety, GPU requirements, numerical behavior".

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@cpp/include/cuopt/mathematical_optimization/remote_solve_registry.hpp` around
lines 25 - 30, Add Doxygen-style documentation immediately before
solve_lp_remote_fn_t and solve_mip_remote_fn_t, describing that callbacks return
ownership via unique_ptr, must not throw across the component boundary, and are
instantiated only for <int, double>. Retain the existing intent comment while
documenting the callback contract for both typedefs.

Source: Path instructions

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@cpp/CMakeLists.txt`:
- Around line 865-890: Extend _UMBRELLA_SRC with used symbol-reference anchors
for exported symbols from cuopt_base and cuopt_routing so their libraries remain
in DT_NEEDED under --as-needed; add the routing anchor only under the
appropriate routing build guard, while preserving the existing LP anchor and
SKIP_GRPC_BUILD gRPC guard.

In `@cpp/cuopt_cli.cpp`:
- Around line 203-205: Replace the unchecked static_cast in the remote execution
path with a validated dynamic_cast to cpu_optimization_problem_t. Before passing
cpu_prob to solve_*_remote, enforce that the cast succeeded and fail cleanly
using the existing cuopt_expects-style validation, preserving the GPU backend
path without reinterpreting its problem type.

In `@cpp/include/cuopt/mathematical_optimization/remote_solve_registry.hpp`:
- Around line 32-38: Synchronize remote-solver registry access by declaring
g_solve_lp_remote_fn and g_solve_mip_remote_fn as atomic function-pointer slots
in cpp/include/cuopt/mathematical_optimization/remote_solve_registry.hpp, and
store registrations with release ordering in
cpp/src/pdlp/remote_solve_registry.cpp. In cpp/src/pdlp/solve.cu lines
2696-2700, load g_solve_lp_remote_fn once into a local before the null check and
reuse it for dispatch; apply the same pattern to g_solve_mip_remote_fn in
cpp/src/mip_heuristics/solve.cu lines 918-922.
- Around line 19-23: Change the forward declarations of pdlp_solver_settings_t
and mip_solver_settings_t to use class instead of struct, matching their class
definitions and keeping both template parameter lists unchanged.

In `@python/libcuopt/CMakeLists.txt`:
- Around line 98-102: Update the RPATH property list in the CMake configuration
to include the cuopt_routing target with the same ${rpaths} append used by
cuopt_base and cuopt_lp, ensuring its installed shared library receives the
bundled CUDA library runtime paths.

---

Nitpick comments:
In `@cpp/include/cuopt/mathematical_optimization/remote_solve_registry.hpp`:
- Around line 25-30: Add Doxygen-style documentation immediately before
solve_lp_remote_fn_t and solve_mip_remote_fn_t, describing that callbacks return
ownership via unique_ptr, must not throw across the component boundary, and are
instantiated only for <int, double>. Retain the existing intent comment while
documenting the callback contract for both typedefs.

In `@cpp/src/mip_heuristics/solve.cu`:
- Around line 918-922: Replace the inline lazy-load and null-check sequence in
the MIP remote execution path with a shared registry helper such as
ensure_remote_solvers_loaded(). Move loader-name handling, dlopen result
validation, dlerror reporting, and synchronization into that helper, then reuse
it from both this path and the corresponding LP path while preserving the
existing runtime-error behavior when the remote solver remains unavailable.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: ea594003-fad8-4a06-acda-c89669a7e541

📥 Commits

Reviewing files that changed from the base of the PR and between ec81ebc and 80c1c34.

📒 Files selected for processing (16)

• ci/build_wheel_cuopt.sh
• ci/test_skills_assets.sh
• conda/recipes/libcuopt/recipe.yaml
• cpp/CMakeLists.txt
• cpp/cuopt_cli.cpp
• cpp/include/cuopt/mathematical_optimization/remote_solve_registry.hpp
• cpp/src/grpc/client/grpc_registration.cpp
• cpp/src/mip_heuristics/solve.cu
• cpp/src/pdlp/CMakeLists.txt
• cpp/src/pdlp/remote_solve_registry.cpp
• cpp/src/pdlp/solve.cu
• cpp/src/routing/CMakeLists.txt
• docs/cuopt/source/conf.py
• docs/cuopt/source/cuopt-c/convex/examples/Makefile
• docs/cuopt/source/cuopt-c/mip/examples/Makefile
• python/libcuopt/CMakeLists.txt

🚧 Files skipped from review as they are similar to previous changes (1)

• cpp/src/routing/CMakeLists.txt

[coderabbitai]

🩺 Stability & Availability | 🔴 Critical | 🏗️ Heavy lift

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
rg -n 'SKIP_GRPC_BUILD|CUOPT_ENABLE_GRPC' cpp/CMakeLists.txt
rg -n '_cuopt_base_anchor|_cuopt_routing_anchor|_UMBRELLA_SRC' cpp/CMakeLists.txt

Repository: NVIDIA/cuopt

Length of output: 1129

---

🏁 Script executed:

#!/bin/bash
set -eu
printf '%s\n' '--- CMake target/link context ---'
sed -n '520,575p;780,850p;850,935p;1050,1100p' cpp/CMakeLists.txt
printf '%s\n' '--- component target and exported-symbol references ---'
rg -n 'add_library|target_sources|target_link_libraries|install|cuopt_base|cuopt_routing|cuOptDestroyProblem|solve_lp_remote' cpp/CMakeLists.txt cpp/src cpp/include -g '*.{cmake,txt,cpp,cu,c,h,hpp}' | head -300
printf '%s\n' '--- candidate component files ---'
git ls-files | rg '(^|/)(base|routing|.*base.*|.*routing.*)|solve_remote|remote_solve_registry'

Repository: NVIDIA/cuopt

Length of output: 32659

---

🏁 Script executed:

#!/bin/bash
set -eu
printf '%s\n' '--- option resolution and gRPC target conditions ---'
sed -n '35,65p;350,410p;610,715p;930,1010p;1150,1230p' cpp/CMakeLists.txt
printf '%s\n' '--- linker/as-needed configuration ---'
rg -n --hidden -i 'as-needed|no-as-needed|copy-dt-needed|link_options|CMAKE_SHARED_LINKER_FLAGS|CMAKE_EXE_LINKER_FLAGS' . -g '!**/build/**' -g '!**/node_modules/**'
printf '%s\n' '--- public base/routing declarations and definitions ---'
rg -n '^(export )?(namespace|class|struct)|(^|[^A-Za-z0-9_])(solve|cuOpt[A-Z]|cuopt_[A-Za-z0-9_]+)\s*\\(' cpp/include/cuopt/routing cpp/include/cuopt/mathematical_optimization cpp/src/base cpp/src/routing cpp/src/pdlp -g '*.{h,hpp,cpp,cu}' | head -300

Repository: NVIDIA/cuopt

Length of output: 16472

---

🏁 Script executed:

#!/bin/bash
set -eu
printf '%s\n' '--- component source lists ---'
sed -n '410,525p' cpp/CMakeLists.txt
printf '%s\n' '--- installed/public routing API declarations ---'
rg -n 'namespace cuopt|class |struct |cuOpt|solve\\s*\\(' cpp/include/cuopt/routing cpp/include/cuopt/mathematical_optimization -g '*.{h,hpp}' | head -240
printf '%s\n' '--- routing/base definitions and exported C API definitions ---'
rg -n 'namespace cuopt|cuOpt[A-Z][A-Za-z0-9_]*\\s*\\(|^[A-Za-z_][A-Za-z0-9_:<>*& ]*::[A-Za-z_][A-Za-z0-9_]*\\s*\\(' cpp/src/routing cpp/src/base cpp/src/pdlp -g '*.{cpp,cu}' | head -240
printf '%s\n' '--- all CUOPT_ENABLE_GRPC references ---'
rg -n --hidden 'CUOPT_ENABLE_GRPC|SKIP_GRPC_BUILD' . -g '!**/build/**' -g '!**/node_modules/**'

Repository: NVIDIA/cuopt

Length of output: 9218

---

🏁 Script executed:

#!/bin/bash
python3 - <<'PY'
from pathlib import Path
import re

text = Path("cpp/CMakeLists.txt").read_text()

umbrella = re.search(
    r"set\(_UMBRELLA_SRC \[=\[(.*?)\]=\]\)",
    text,
    re.S,
).group(1)
anchors = {
    name: name in umbrella
    for name in ("_cuopt_lp_anchor", "_cuopt_grpc_anchor",
                 "_cuopt_base_anchor", "_cuopt_routing_anchor")
}

cuopt_links = re.search(
    r"if\(NOT SKIP_ROUTING_BUILD\)\s*"
    r"target_link_libraries\(cuopt PUBLIC (.*?)\)\s*"
    r"else\(\)\s*"
    r"target_link_libraries\(cuopt PUBLIC (.*?)\)",
    text,
    re.S,
)
grpc_link = re.search(
    r"if\(NOT SKIP_GRPC_BUILD\)\s*"
    r"target_link_libraries\(cuopt PUBLIC cuopt_grpc\)",
    text,
)

print("umbrella anchors:", anchors)
print("routing-enabled umbrella dependencies:",
      " ".join(cuopt_links.group(1).split()) if cuopt_links else "<not found>")
print("routing-disabled umbrella dependencies:",
      " ".join(cuopt_links.group(2).split()) if cuopt_links else "<not found>")
print("gRPC umbrella dependency guarded by SKIP_GRPC_BUILD:",
      bool(grpc_link))
print("gRPC anchor guarded by SKIP_GRPC_BUILD:",
      "if(NOT SKIP_GRPC_BUILD)" in text[text.index("if(NOT SKIP_GRPC_BUILD)", text.index("set(_UMBRELLA_SRC")):])

assert anchors["_cuopt_lp_anchor"]
assert not anchors["_cuopt_base_anchor"]
assert not anchors["_cuopt_routing_anchor"]
assert cuopt_links and "cuopt_base" in cuopt_links.group(1)
assert cuopt_links and "cuopt_routing" in cuopt_links.group(1)
assert grpc_link
PY

Repository: NVIDIA/cuopt

Length of output: 521

---

Add real DT_NEEDED anchors for cuopt_base and cuopt_routing.

cuopt links these components publicly, but the umbrella TU references only LP and gRPC. With --as-needed, base and routing can be omitted, breaking consumers that use -lcuopt. Add exported-symbol anchors, conditionally for routing. The gRPC guards are already aligned with SKIP_GRPC_BUILD.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@cpp/CMakeLists.txt` around lines 865 - 890, Extend _UMBRELLA_SRC with used
symbol-reference anchors for exported symbols from cuopt_base and cuopt_routing
so their libraries remain in DT_NEEDED under --as-needed; add the routing anchor
only under the appropriate routing build guard, while preserving the existing LP
anchor and SKIP_GRPC_BUILD gRPC guard.

Source: Path instructions

[coderabbitai]

🩺 Stability & Availability | 🔴 Critical | ⚡ Quick win

Unchecked static_cast to cpu_optimization_problem_t — wrong downcast when remote execution is enabled with the GPU backend.

is_remote_execution_enabled() and get_memory_backend_type() are independent conditions. If the backend resolves to GPU (Lines 136-140 construct an optimization_problem_t) while CUOPT_REMOTE_HOST/CUOPT_REMOTE_PORT are set, this static_cast reinterprets a GPU problem as a CPU problem and *cpu_prob is then handed to solve_*_remote — undefined behavior, most likely a corrupt read rather than a clean error.

The library-side dispatchers guard this exact conversion (cpp/src/pdlp/solve.cu Line 2692 and cpp/src/mip_heuristics/solve.cu Line 914 use dynamic_cast plus cuopt_expects); the CLI should not be the one place that skips the check.

🛡️ Proposed fix: validate the downcast

       auto* cpu_prob =
-        static_cast<cuopt::mathematical_optimization::cpu_optimization_problem_t<int, double>*>(
+        dynamic_cast<cuopt::mathematical_optimization::cpu_optimization_problem_t<int, double>*>(
           problem_interface.get());
+      if (cpu_prob == nullptr) {
+        CUOPT_LOG_ERROR(
+          "Remote execution requires the CPU memory backend, but a GPU-backed problem was "
+          "created. Unset CUOPT_REMOTE_HOST/CUOPT_REMOTE_PORT or select the CPU backend.");
+        return -1;
+      }
       if (is_mip) {

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	auto* cpu_prob =
	static_cast<cuopt::mathematical_optimization::cpu_optimization_problem_t<int, double>*>(
	problem_interface.get());
	auto* cpu_prob =
	dynamic_cast<cuopt::mathematical_optimization::cpu_optimization_problem_t<int, double>*>(
	problem_interface.get());
	if (cpu_prob == nullptr) {
	CUOPT_LOG_ERROR(
	"Remote execution requires the CPU memory backend, but a GPU-backed problem was "
	"created. Unset CUOPT_REMOTE_HOST/CUOPT_REMOTE_PORT or select the CPU backend.");
	return -1;
	}
	if (is_mip) {

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@cpp/cuopt_cli.cpp` around lines 203 - 205, Replace the unchecked static_cast
in the remote execution path with a validated dynamic_cast to
cpu_optimization_problem_t. Before passing cpu_prob to solve_*_remote, enforce
that the cast succeeded and fail cleanly using the existing cuopt_expects-style
validation, preserving the GPU backend path without reinterpreting its problem
type.

[coderabbitai]

📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
rg -nP --type=cpp --type=cuda -g '!**/tests/**' '\b(class|struct)\s+(pdlp_solver_settings_t|mip_solver_settings_t|cpu_optimization_problem_t)\b' | head -50

Repository: NVIDIA/cuopt

Length of output: 150

---

🏁 Script executed:

#!/bin/bash
set -eu
printf '%s\n' 'Definitions and declarations across tracked source files:'
rg -n --glob '!**/tests/**' '\b(class|struct)\s+(pdlp_solver_settings_t|mip_solver_settings_t|cpu_optimization_problem_t)\b' .
printf '%s\n' 'References to the settings types:'
rg -n --glob '!**/tests/**' '\b(pdlp_solver_settings_t|mip_solver_settings_t)\b' cpp python 2>/dev/null | head -100

Repository: NVIDIA/cuopt

Length of output: 14060

---

🏁 Script executed:

#!/bin/bash
set -eu
printf '%s\n' 'remote_solve_registry.hpp:'
sed -n '1,80p' cpp/include/cuopt/mathematical_optimization/remote_solve_registry.hpp
printf '%s\n' 'Definitions:'
sed -n '100,130p' cpp/include/cuopt/mathematical_optimization/pdlp/solver_settings.hpp
sed -n '50,75p' cpp/include/cuopt/mathematical_optimization/mip/solver_settings.hpp
printf '%s\n' 'Include/use context:'
rg -n -C 3 'remote_solve_registry|pdlp_solver_settings_t|mip_solver_settings_t' cpp/include/cuopt/mathematical_optimization/remote_solve_registry.hpp cpp/include/cuopt/mathematical_optimization/solve_remote.hpp cpp/src -g '*.cpp' -g '*.cu' -g '*.hpp' -g '*.cuh' | head -180

Repository: NVIDIA/cuopt

Length of output: 18505

---

🏁 Script executed:

#!/bin/bash
set -eu
if command -v g++ >/dev/null 2>&1; then
  printf '%s\n' 'Compiler behavior for mismatched class-keys:'
  printf '%s\n' 'template<class T> class X; template<class T> struct X {}; int main() {}' |
    g++ -std=c++17 -x c++ -fsyntax-only -Wall -Wextra -
else
  printf '%s\n' 'g++ unavailable; source evidence is sufficient to identify the class-key mismatch.'
fi

Repository: NVIDIA/cuopt

Length of output: 233

---

Use class for both forward declarations.

These templates are defined as classes; the mismatched class-keys are ill-formed when both declarations are visible.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@cpp/include/cuopt/mathematical_optimization/remote_solve_registry.hpp` around
lines 19 - 23, Change the forward declarations of pdlp_solver_settings_t and
mip_solver_settings_t to use class instead of struct, matching their class
definitions and keeping both template parameter lists unchanged.

[coderabbitai]

🩺 Stability & Availability | 🟠 Major | ⚡ Quick win

Unsynchronized access to the remote-solver registry across three sites. The shared root cause is that the registry slots are declared as plain function pointers: they are published from libcuopt_grpc.so's ELF constructor during a lazy dlopen while other threads read them with no synchronization, so concurrent solve_lp/solve_mip calls race on a non-atomic object.

• cpp/include/cuopt/mathematical_optimization/remote_solve_registry.hpp#L32-L38: declare both slots as std::atomic<...> (and store with release ordering in cpp/src/pdlp/solve.cu's sibling remote_solve_registry.cpp); this is the only edit that fixes the race.
• cpp/src/pdlp/solve.cu#L2696-L2700: once the slot is atomic, load it once into a local before the null test and reuse that local for the dispatch, rather than reading the global three times.
• cpp/src/mip_heuristics/solve.cu#L918-L922: apply the same single-load-into-a-local pattern for g_solve_mip_remote_fn.

📍 Affects 3 files

• cpp/include/cuopt/mathematical_optimization/remote_solve_registry.hpp#L32-L38 (this comment)
• cpp/src/pdlp/solve.cu#L2696-L2700
• cpp/src/mip_heuristics/solve.cu#L918-L922

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@cpp/include/cuopt/mathematical_optimization/remote_solve_registry.hpp` around
lines 32 - 38, Synchronize remote-solver registry access by declaring
g_solve_lp_remote_fn and g_solve_mip_remote_fn as atomic function-pointer slots
in cpp/include/cuopt/mathematical_optimization/remote_solve_registry.hpp, and
store registrations with release ordering in
cpp/src/pdlp/remote_solve_registry.cpp. In cpp/src/pdlp/solve.cu lines
2696-2700, load g_solve_lp_remote_fn once into a local before the null check and
reuse it for dispatch; apply the same pattern to g_solve_mip_remote_fn in
cpp/src/mip_heuristics/solve.cu lines 918-922.

[coderabbitai]

🩺 Stability & Availability | 🟠 Major | ⚡ Quick win

cuopt_routing is missing from the RPATH list.

cuopt_base and cuopt_lp were given the same ${rpaths} append as cuopt, cuopt_cli, and cuopt_grpc_server, but cuopt_routing (introduced as its own component library in this PR) isn't included. If it links any of the bundled CUDA libs the RPATH set targets (cublas/cusparse/cusolver/nccl/nvjitlink), the wheel-installed libcuopt_routing.so could fail to locate them at runtime.

🔧 Suggested fix

 set_property(TARGET cuopt PROPERTY INSTALL_RPATH ${rpaths} APPEND)
 set_property(TARGET cuopt_base PROPERTY INSTALL_RPATH ${rpaths} APPEND)
+set_property(TARGET cuopt_routing PROPERTY INSTALL_RPATH ${rpaths} APPEND)
 set_property(TARGET cuopt_lp PROPERTY INSTALL_RPATH ${rpaths} APPEND)
 set_property(TARGET cuopt_cli PROPERTY INSTALL_RPATH ${rpaths} APPEND)
 set_property(TARGET cuopt_grpc_server PROPERTY INSTALL_RPATH ${rpaths} APPEND)

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	set_property(TARGET cuopt PROPERTY INSTALL_RPATH ${rpaths} APPEND)
	set_property(TARGET cuopt_base PROPERTY INSTALL_RPATH ${rpaths} APPEND)
	set_property(TARGET cuopt_lp PROPERTY INSTALL_RPATH ${rpaths} APPEND)
	set_property(TARGET cuopt_cli PROPERTY INSTALL_RPATH ${rpaths} APPEND)
	set_property(TARGET cuopt_grpc_server PROPERTY INSTALL_RPATH ${rpaths} APPEND)
	set_property(TARGET cuopt PROPERTY INSTALL_RPATH ${rpaths} APPEND)
	set_property(TARGET cuopt_base PROPERTY INSTALL_RPATH ${rpaths} APPEND)
	set_property(TARGET cuopt_routing PROPERTY INSTALL_RPATH ${rpaths} APPEND)
	set_property(TARGET cuopt_lp PROPERTY INSTALL_RPATH ${rpaths} APPEND)
	set_property(TARGET cuopt_cli PROPERTY INSTALL_RPATH ${rpaths} APPEND)
	set_property(TARGET cuopt_grpc_server PROPERTY INSTALL_RPATH ${rpaths} APPEND)

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@python/libcuopt/CMakeLists.txt` around lines 98 - 102, Update the RPATH
property list in the CMake configuration to include the cuopt_routing target
with the same ${rpaths} append used by cuopt_base and cuopt_lp, ensuring its
installed shared library receives the bundled CUDA library runtime paths.