cpp/cuopt_cli.cpp (1)

69-72: Dead code: make_async() is now unused.

The make_async() helper function at line 71 is no longer called anywhere in this file after the migration to value-type memory resources. Consider removing it.

🧹 Proposed removal

-/**
- * `@brief` Make an async memory resource for RMM
- * `@return` std::shared_ptr<rmm::mr::cuda_async_memory_resource>
- */
-inline auto make_async() { return std::make_shared<rmm::mr::cuda_async_memory_resource>(); }
-

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@cpp/cuopt_cli.cpp` around lines 69 - 72, The helper function make_async() is
now unused and should be removed: delete the inline auto make_async() { return
std::make_shared<rmm::mr::cuda_async_memory_resource>(); }
declaration/definition to eliminate dead code and any associated includes or
comments that only served it; ensure no remaining references to make_async exist
and run a build to confirm no usages remain.

cpp/src/pdlp/termination_strategy/infeasibility_information.cu (1)

18-19: Remove the duplicate Thrust include.

Line 18 adds <thrust/iterator/transform_output_iterator.h>, which is already included at Line 32. Keeping only one include improves clarity.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@cpp/src/pdlp/termination_strategy/infeasibility_information.cu` around lines
18 - 19, Remove the duplicate Thrust include by deleting the redundant `#include`
for <thrust/iterator/transform_output_iterator.h> so only the single existing
include remains; locate the duplicate include line (the
<thrust/iterator/transform_output_iterator.h> directive) in
infeasibility_information.cu and remove it, leaving the original include at the
later occurrence intact.

Fix all unresolved CodeRabbit comments on this PR:

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

cpp/src/utilities/cuda_helpers.cuh (1)

241-246: ⚠️ Potential issue | 🟠 Major

Add cuopt-native CUDA error checking for cudaMemGetInfo.

At Line 244, cudaMemGetInfo is unchecked. On failure, total_mem can be invalid and the CUDA error state handling is inconsistent. Please wrap it with the cuopt-native CUDA check macro (not RAFT).

Suggested patch

 inline size_t get_device_memory_size()
 {
-  size_t free_mem, total_mem;
-  cudaMemGetInfo(&free_mem, &total_mem);
+  size_t free_mem{0}, total_mem{0};
+  CUDA_CHECK(cudaMemGetInfo(&free_mem, &total_mem));  // or cuopt-native equivalent from utilities/macros.cuh
   // TODO (bdice): Restore limiting adaptor check after updating CCCL to support resource_cast
   return total_mem;
 }

#!/bin/bash
# Verify the exact cuopt-native CUDA check macro and current usage patterns.

set -euo pipefail

echo "== locate macros header =="
fd -i 'macros.cuh'

echo "== CUDA-related macro definitions in macros header(s) =="
fd -i 'macros.cuh' | xargs -r rg -n 'define\s+.*(CUDA|cuda|CHECK)'

echo "== current get_device_memory_size implementation =="
rg -n -C3 'get_device_memory_size|cudaMemGetInfo' cpp/src/utilities/cuda_helpers.cuh

echo "== existing CUDA check macro usage in utilities =="
rg -n 'CUDA_CHECK|CHECK_CUDA|CUOPT_.*CUDA|cudaGetLastError' cpp/src/utilities

As per coding guidelines: "**/*.{cu,cuh}: Every CUDA kernel launch and memory operation must have error checking with CUDA_CHECK or equivalent verification." Based on learnings: "cuopt code should use cuopt-native error-checking macros instead of depending on RAFT-specific ones."

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@cpp/src/utilities/cuda_helpers.cuh` around lines 241 - 246, The
cudaMemGetInfo call inside get_device_memory_size lacks cuopt-native error
checking; wrap the cudaMemGetInfo(&free_mem, &total_mem) invocation with the
project’s cuopt-native CUDA-check macro (e.g., CUDA_CHECK or the equivalent
cuopt macro used across the codebase) so errors are asserted/handled
consistently (preserve the free_mem/total_mem variables and return total_mem as
before), and ensure the header that defines the macro is included if not
already.

ci/build_wheel_libcuopt.sh (1)

8-49: ⚠️ Potential issue | 🟠 Major

Sourced PR constraints are likely discarded by later PIP_CONSTRAINT reassignment.

After Line 8 sources use_wheels_from_prs.sh, Line 48 points PIP_CONSTRAINT to /tmp/constraints.txt, which can bypass previously appended entries.

Minimal fix

 CUOPT_MPS_PARSER_WHEELHOUSE=$(RAPIDS_PY_WHEEL_NAME="cuopt_mps_parser" rapids-download-wheels-from-github python)
-echo "cuopt-mps-parser @ file://$(echo ${CUOPT_MPS_PARSER_WHEELHOUSE}/cuopt_mps_parser*.whl)" >> /tmp/constraints.txt
-export PIP_CONSTRAINT="/tmp/constraints.txt"
+echo "cuopt-mps-parser @ file://$(echo "${CUOPT_MPS_PARSER_WHEELHOUSE}"/cuopt_mps_parser*.whl)" >> "${PIP_CONSTRAINT}"

#!/bin/bash
set -euo pipefail
rg -n 'use_wheels_from_prs\.sh|/tmp/constraints\.txt|export PIP_CONSTRAINT=' ci/build_wheel_libcuopt.sh

Expected: helper sourcing appears before reassignment to /tmp/constraints.txt.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@ci/build_wheel_libcuopt.sh` around lines 8 - 49, The script currently
overwrites any constraints set by sourcing use_wheels_from_prs.sh because it
later unconditionally exports PIP_CONSTRAINT="/tmp/constraints.txt"; fix by
preserving or merging prior constraints: either move the source
./ci/use_wheels_from_prs.sh line to after the block that writes
/tmp/constraints.txt (so the PR helper runs last) or modify the logic around
CUOPT_MPS_PARSER_WHEELHOUSE and the export PIP_CONSTRAINT to append to an
existing PIP_CONSTRAINT if present (check $PIP_CONSTRAINT and include it when
writing /tmp/constraints.txt or export a combined value); update references in
this script to PIP_CONSTRAINT and CUOPT_MPS_PARSER_WHEELHOUSE accordingly.

ci/test_wheel_cuopt.sh (1)

11-27: ⚠️ Potential issue | 🟠 Major

PIP_CONSTRAINT entries from the sourced helper are being clobbered.

Line 11 appends PR wheel pins via use_wheels_from_prs.sh, but Line 22 recreates the file with cat >, dropping those earlier constraints.

Minimal fix

-cat > "${PIP_CONSTRAINT}" <<EOF
+cat >> "${PIP_CONSTRAINT}" <<EOF
 cuopt-${RAPIDS_PY_CUDA_SUFFIX} @ file://$(echo ${CUOPT_WHEELHOUSE}/cuopt_${RAPIDS_PY_CUDA_SUFFIX}-*.whl)
 cuopt-mps-parser @ file://$(echo ${CUOPT_MPS_PARSER_WHEELHOUSE}/cuopt_mps_parser-*.whl)
 cuopt-sh-client @ file://$(echo ${CUOPT_SH_CLIENT_WHEELHOUSE}/cuopt_sh_client-*.whl)
 libcuopt-${RAPIDS_PY_CUDA_SUFFIX} @ file://$(echo ${LIBCUOPT_WHEELHOUSE}/libcuopt_${RAPIDS_PY_CUDA_SUFFIX}-*.whl)
 EOF

#!/bin/bash
set -euo pipefail
rg -n 'use_wheels_from_prs\.sh|cat > "\$\{PIP_CONSTRAINT\}"|cat >> "\$\{PIP_CONSTRAINT\}"' ci/test_wheel_cuopt.sh

Expected: shows helper sourcing plus a truncating cat > (current behavior).

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@ci/test_wheel_cuopt.sh` around lines 11 - 27, The PIP_CONSTRAINT file is
being overwritten after sourcing use_wheels_from_prs.sh because the here-doc
uses "cat > \"${PIP_CONSTRAINT}\"", which drops entries added by the helper;
change the redirection to append (use "cat >> \"${PIP_CONSTRAINT}\"") so the
block that writes cuopt wheel pins (the here-doc that writes
cuopt-${RAPIDS_PY_CUDA_SUFFIX}, cuopt-mps-parser, cuopt-sh-client,
libcuopt-${RAPIDS_PY_CUDA_SUFFIX}) is appended to the existing ${PIP_CONSTRAINT}
written/modified by use_wheels_from_prs.sh instead of truncating it.