Add a build.sh flag to gate multi-GPU PDLP deps and build

build.sh

@@ -14,7 +14,7 @@ ARGS=$*
 REPODIR=$(cd "$(dirname "$0")"; pwd)
 LIBCUOPT_BUILD_DIR=${LIBCUOPT_BUILD_DIR:=${REPODIR}/cpp/build}
 
-VALIDARGS="clean codegen libcuopt cuopt_grpc_server cuopt cuopt_server cuopt_sh_client docs deb -a -b -g -fsanitize -tsan -msan -v -l= --verbose-pdlp --build-lp-only  --no-fetch-rapids --skip-c-python-adapters --skip-tests-build --skip-routing-build --skip-grpc-build --skip-fatbin-write --host-lineinfo --split-compile [--cmake-args=\\\"<args>\\\"] [--cache-tool=<tool>] --install --allgpuarch --ci-only-arch --show_depr_warn -h --help"
+VALIDARGS="clean codegen libcuopt cuopt_grpc_server cuopt cuopt_server cuopt_sh_client docs deb -a -b -g -fsanitize -tsan -msan -v -l= --verbose-pdlp --build-lp-only  --no-fetch-rapids --skip-c-python-adapters --skip-tests-build --skip-routing-build --skip-grpc-build --skip-distributed-pdlp-build --skip-fatbin-write --host-lineinfo --split-compile [--cmake-args=\\\"<args>\\\"] [--cache-tool=<tool>] --install --allgpuarch --ci-only-arch --show_depr_warn -h --help"
 HELP="$0 [<target> ...] [<flag> ...]
  where <target> is:
    clean            - remove all existing build artifacts and configuration (start over)
@@ -43,6 +43,7 @@ HELP="$0 [<target> ...] [<flag> ...]
    --skip-tests-build  - disable building of all tests
    --skip-routing-build - skip building routing components
    --skip-grpc-build    - skip building gRPC and protobuf components (auto-enabled with -tsan)
+   --skip-distributed-pdlp-build - skip distributed PDLP, NCCL, and KaMinPar
    --skip-fatbin-write      - skip the fatbin write
    --host-lineinfo           - build with debug line information for host code
    --split-compile           - opt in to nvcc split compilation; builds may be nondeterministic
@@ -84,6 +85,7 @@ SKIP_C_PYTHON_ADAPTERS=0
 SKIP_TESTS_BUILD=0
 SKIP_ROUTING_BUILD=0
 SKIP_GRPC_BUILD=0
+SKIP_DISTRIBUTED_PDLP_BUILD=0
 WRITE_FATBIN=1
 HOST_LINEINFO=0
 CACHE_ARGS=()
@@ -255,6 +257,9 @@ fi
 if hasArg --skip-grpc-build; then
     SKIP_GRPC_BUILD=1
 fi
+if hasArg --skip-distributed-pdlp-build; then
+    SKIP_DISTRIBUTED_PDLP_BUILD=1
+fi
 if hasArg --skip-fatbin-write; then
     WRITE_FATBIN=0
 fi
@@ -399,6 +404,7 @@ if buildAll || hasArg libcuopt || hasArg cuopt_grpc_server; then
           -DBUILD_TESTS=$((1 - ${SKIP_TESTS_BUILD})) \
           -DSKIP_ROUTING_BUILD=${SKIP_ROUTING_BUILD} \
           -DSKIP_GRPC_BUILD=${SKIP_GRPC_BUILD} \
+          -DSKIP_DISTRIBUTED_PDLP_BUILD=${SKIP_DISTRIBUTED_PDLP_BUILD} \
           -DWRITE_FATBIN=${WRITE_FATBIN} \
           -DHOST_LINEINFO=${HOST_LINEINFO} \
           -DPARALLEL_LEVEL="${PARALLEL_LEVEL}" \

cpp/CMakeLists.txt

@@ -52,6 +52,7 @@ option(BUILD_LP_ONLY "Build only linear programming components, exclude routing
 option(SKIP_C_PYTHON_ADAPTERS "Skip building C and Python adapter files (cython_solve.cu and cuopt_c.cpp)" OFF)
 option(SKIP_ROUTING_BUILD "Skip building routing components" OFF)
 option(SKIP_GRPC_BUILD "Skip building gRPC and protobuf components" OFF)
+option(SKIP_DISTRIBUTED_PDLP_BUILD "Skip distributed PDLP, NCCL, and KaMinPar" OFF)
 option(WRITE_FATBIN "Enable fatbin writing" ON)
 option(HOST_LINEINFO "Build with debug line information for host code" OFF)
 
@@ -62,9 +63,14 @@ message(VERBOSE "cuOpt: Disable OpenMP: ${DISABLE_OPENMP}")
 message(VERBOSE "cuOpt: Build LP-only mode: ${BUILD_LP_ONLY}")
 message(VERBOSE "cuOpt: Skip C/Python adapters: ${SKIP_C_PYTHON_ADAPTERS}")
 message(VERBOSE "cuOpt: Skip routing build: ${SKIP_ROUTING_BUILD}")
+message(VERBOSE "cuOpt: Skip distributed PDLP build: ${SKIP_DISTRIBUTED_PDLP_BUILD}")
 message(VERBOSE "cuOpt: Build with debug line information for host code: ${HOST_LINEINFO}")
 message(VERBOSE "cuOpt: fatbin: ${WRITE_FATBIN}")
 
+if (NOT SKIP_DISTRIBUTED_PDLP_BUILD)
+    add_compile_definitions(CUOPT_ENABLE_DISTRIBUTED_PDLP)
+endif ()
+
 # ##################################################################################################
 # - compiler options ------------------------------------------------------------------------------
 
@@ -333,39 +339,41 @@ find_package(CUDSS REQUIRED)
 # - NCCL (multi-GPU distributed PDLP) -------------------------------------------------------------
 # NCCL is shipped via the conda env; no canonical CMake config target, so look it
 # up by name in the standard lib paths (plus CONDA_PREFIX as a hint).
-set(NCCL_HINT_PREFIXES "")
-if (DEFINED ENV{CONDA_PREFIX} AND NOT "$ENV{CONDA_PREFIX}" STREQUAL "")
-    list(APPEND NCCL_HINT_PREFIXES "$ENV{CONDA_PREFIX}")
-endif ()
-find_path(NCCL_INCLUDE_DIR
-    NAMES nccl.h
-    HINTS ${NCCL_HINT_PREFIXES}
-    PATH_SUFFIXES include
-)
-find_library(NCCL_LIBRARY
-    NAMES nccl
-    HINTS ${NCCL_HINT_PREFIXES}
-    PATH_SUFFIXES lib lib64
-)
-if (NOT NCCL_INCLUDE_DIR OR NOT NCCL_LIBRARY)
-    message(FATAL_ERROR "NCCL not found. Looked in ${NCCL_HINT_PREFIXES}. Install nccl-dev / libnccl-dev in the active env.")
-endif ()
-add_library(nccl_external UNKNOWN IMPORTED GLOBAL)
-set_target_properties(nccl_external PROPERTIES
-    IMPORTED_LOCATION "${NCCL_LIBRARY}"
-    INTERFACE_INCLUDE_DIRECTORIES "${NCCL_INCLUDE_DIR}"
-)
-message(STATUS "Using NCCL: ${NCCL_LIBRARY}")
-
-# ##################################################################################################
-# - KaMinPar (multi-threaded partitioning for distributed PDLP) ------------------------------------
-# Brought in the RAPIDS way (rapids_cpm_find): uses an installed KaMinPar (deb/rpm/conda,
-# discovered via its CMake config) if present, otherwise builds the pinned source via CPM.
-include(cmake/thirdparty/get_kaminpar.cmake)
-if (NOT TARGET KaMinPar::KaMinPar)
-    message(FATAL_ERROR "KaMinPar::KaMinPar was not made available by get_kaminpar.cmake")
+if (NOT SKIP_DISTRIBUTED_PDLP_BUILD)
+    set(NCCL_HINT_PREFIXES "")
+    if (DEFINED ENV{CONDA_PREFIX} AND NOT "$ENV{CONDA_PREFIX}" STREQUAL "")
+        list(APPEND NCCL_HINT_PREFIXES "$ENV{CONDA_PREFIX}")
+    endif ()
+    find_path(NCCL_INCLUDE_DIR
+        NAMES nccl.h
+        HINTS ${NCCL_HINT_PREFIXES}
+        PATH_SUFFIXES include
+    )
+    find_library(NCCL_LIBRARY
+        NAMES nccl
+        HINTS ${NCCL_HINT_PREFIXES}
+        PATH_SUFFIXES lib lib64
+    )
+    if (NOT NCCL_INCLUDE_DIR OR NOT NCCL_LIBRARY)
+        message(FATAL_ERROR "NCCL not found. Looked in ${NCCL_HINT_PREFIXES}. Install nccl-dev / libnccl-dev in the active env.")
+    endif ()
+    add_library(nccl_external UNKNOWN IMPORTED GLOBAL)
+    set_target_properties(nccl_external PROPERTIES
+        IMPORTED_LOCATION "${NCCL_LIBRARY}"
+        INTERFACE_INCLUDE_DIRECTORIES "${NCCL_INCLUDE_DIR}"
+    )
+    message(STATUS "Using NCCL: ${NCCL_LIBRARY}")
+
+    # ##################################################################################################
+    # - KaMinPar (multi-threaded partitioning for distributed PDLP) ------------------------------------
+    # Brought in the RAPIDS way (rapids_cpm_find): uses an installed KaMinPar (deb/rpm/conda,
+    # discovered via its CMake config) if present, otherwise builds the pinned source via CPM.
+    include(cmake/thirdparty/get_kaminpar.cmake)
+    if (NOT TARGET KaMinPar::KaMinPar)
+        message(FATAL_ERROR "KaMinPar::KaMinPar was not made available by get_kaminpar.cmake")
+    endif ()
+    message(STATUS "Using KaMinPar (distributed PDLP graph partitioner)")
 endif ()
-message(STATUS "Using KaMinPar (distributed PDLP graph partitioner)")
 
 # ##################################################################################################
 # - gRPC and Protobuf setup -----------------------------------------------------------------------
@@ -698,8 +706,6 @@ add_dependencies(cuopt_objs PSLP)
 # "target KaMinPar is not in any export set"). libKaMinPar.a is self-contained (the
 # kaminpar-common OBJECT lib is archived into it); we only need its public headers
 # (<kaminpar.h>, which pulls in stdlib + TBB) at compile time.
-target_include_directories(cuopt_objs SYSTEM PRIVATE
-        $<TARGET_PROPERTY:KaMinPar::KaMinPar,INTERFACE_INCLUDE_DIRECTORIES>)
 # partitioner.cpp includes <kaminpar.h>. Because KaMinPar is linked by file, cuopt
 # does NOT inherit KaMinPar's PUBLIC compile definitions, so we must replicate the ones that
 # affect what <kaminpar.h> compiles to:
@@ -709,11 +715,15 @@ target_include_directories(cuopt_objs SYSTEM PRIVATE
 #                                   with the KAMINPAR_64BIT_* options in get_kaminpar.cmake;
 #                                   otherwise the public API types (e.g. copy_graph's span<>
 #                                   widths) mismatch the prebuilt archive => undefined refs.
-target_compile_definitions(cuopt_objs PRIVATE TBB_PREVIEW_GLOBAL_CONTROL KAMINPAR_64BIT_EDGE_IDS)
 # Real (non-imported) target only exists when KaMinPar is built from source via CPM;
 # force it to build before the objects that include its headers since it is EXCLUDE_FROM_ALL.
-if (TARGET KaMinPar)
+if (NOT SKIP_DISTRIBUTED_PDLP_BUILD)
+    target_include_directories(cuopt_objs SYSTEM PRIVATE
+            $<TARGET_PROPERTY:KaMinPar::KaMinPar,INTERFACE_INCLUDE_DIRECTORIES>)
+    target_compile_definitions(cuopt_objs PRIVATE TBB_PREVIEW_GLOBAL_CONTROL KAMINPAR_64BIT_EDGE_IDS)
+    if (TARGET KaMinPar)
         add_dependencies(cuopt_objs KaMinPar)
+    endif ()
 endif ()
 
 # ##################################################################################################
@@ -790,7 +800,6 @@ if (BUILD_TESTS)
             OpenMP::OpenMP_CXX
             PRIVATE
             ${CUOPT_PRIVATE_CUDA_LIBS}
-            nccl_external
             $<$<BOOL:${CUOPT_ENABLE_GRPC}>:protobuf::libprotobuf>
             $<$<BOOL:${CUOPT_ENABLE_GRPC}>:gRPC::grpc++>
     )
@@ -813,9 +822,11 @@ if (BUILD_TESTS)
     )
     target_link_libraries(cuopt_static PRIVATE $<TARGET_FILE:PSLP>)
     add_dependencies(cuopt_static PSLP)
-    target_link_libraries(cuopt_static PRIVATE $<TARGET_FILE:KaMinPar::KaMinPar>)
-    if (TARGET KaMinPar)
-        add_dependencies(cuopt_static KaMinPar)
+    if (NOT SKIP_DISTRIBUTED_PDLP_BUILD)
+        target_link_libraries(cuopt_static PRIVATE nccl_external $<TARGET_FILE:KaMinPar::KaMinPar>)
+        if (TARGET KaMinPar)
+            add_dependencies(cuopt_static KaMinPar)
+        endif ()
     endif ()
     add_subdirectory(tests)
 endif (BUILD_TESTS)
@@ -848,7 +859,6 @@ target_link_libraries(cuopt
         ${CUDSS_LIB_FILE}
         PRIVATE
         ${CUOPT_PRIVATE_CUDA_LIBS}
-        nccl_external
         simde::simde
         OpenMP::OpenMP_CXX
         OpenMP::OpenMP_CUDA
@@ -857,9 +867,11 @@ target_link_libraries(cuopt
 )
 target_link_libraries(cuopt PRIVATE $<TARGET_FILE:PSLP>)
 add_dependencies(cuopt PSLP)
-target_link_libraries(cuopt PRIVATE $<TARGET_FILE:KaMinPar::KaMinPar>)
-if (TARGET KaMinPar)
+if (NOT SKIP_DISTRIBUTED_PDLP_BUILD)
+    target_link_libraries(cuopt PRIVATE nccl_external $<TARGET_FILE:KaMinPar::KaMinPar>)
+    if (TARGET KaMinPar)
         add_dependencies(cuopt KaMinPar)
+    endif ()
 endif ()
 # Propagate compile definitions that consumers need when including cuopt headers.
 # These were on cuopt directly before the cuopt_objs refactor; $<TARGET_OBJECTS:...>

cpp/src/pdlp/CMakeLists.txt

@@ -29,13 +29,18 @@ set(LP_CORE_FILES
   ${CMAKE_CURRENT_SOURCE_DIR}/termination_strategy/convergence_information.cu
   ${CMAKE_CURRENT_SOURCE_DIR}/optimal_batch_size_handler/optimal_batch_size_handler.cu
   ${CMAKE_CURRENT_SOURCE_DIR}/utilities/ping_pong_graph.cu
-  ${CMAKE_CURRENT_SOURCE_DIR}/distributed_pdlp/shard.cu
-  ${CMAKE_CURRENT_SOURCE_DIR}/distributed_pdlp/multi_gpu_engine.cu
-  ${CMAKE_CURRENT_SOURCE_DIR}/distributed_pdlp/distributed_algorithms.cu
-  ${CMAKE_CURRENT_SOURCE_DIR}/distributed_pdlp/distributed_utils.cu
-  ${CMAKE_CURRENT_SOURCE_DIR}/distributed_pdlp/partitioner.cpp
 )
 
+if(NOT SKIP_DISTRIBUTED_PDLP_BUILD)
+  list(APPEND LP_CORE_FILES
+    ${CMAKE_CURRENT_SOURCE_DIR}/distributed_pdlp/shard.cu
+    ${CMAKE_CURRENT_SOURCE_DIR}/distributed_pdlp/multi_gpu_engine.cu
+    ${CMAKE_CURRENT_SOURCE_DIR}/distributed_pdlp/distributed_algorithms.cu
+    ${CMAKE_CURRENT_SOURCE_DIR}/distributed_pdlp/distributed_utils.cu
+    ${CMAKE_CURRENT_SOURCE_DIR}/distributed_pdlp/partitioner.cpp
+  )
+endif()
+
 # C and Python adapter files
 set(LP_ADAPTER_FILES
   ${CMAKE_CURRENT_SOURCE_DIR}/utilities/cython_solve.cu

cpp/src/pdlp/pdhg.cu

@@ -4,7 +4,9 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 /* clang-format on */
+#ifdef CUOPT_ENABLE_DISTRIBUTED_PDLP
 #include <pdlp/distributed_pdlp/multi_gpu_engine.hpp>
+#endif
 #include <pdlp/pdhg.hpp>
 #include <pdlp/pdlp.cuh>
 #include <pdlp/pdlp_climber_strategy.hpp>
@@ -510,10 +512,12 @@ void pdhg_solver_t<i_t, f_t>::compute_At_y()
 
   // Multi-GPU dispatch: when the master pdhg has an engine, drive halo
   // exchange + per-shard SpMV via the engine.
+#ifdef CUOPT_ENABLE_DISTRIBUTED_PDLP
   if (is_distributed_master()) {
     mgpu_engine_->distributed_compute_At_y();
     return;
   }
+#endif
 
   if (!batch_mode_) {
     if constexpr (std::is_same_v<f_t, double>) {
@@ -568,10 +572,12 @@ void pdhg_solver_t<i_t, f_t>::compute_A_x()
   // Multi-GPU dispatch: see compute_At_y. The engine halo-updates the
   // reflected_primal vector (the buffer this SpMV reads) and then drives
   // per-shard local cusparse SpMV.
+#ifdef CUOPT_ENABLE_DISTRIBUTED_PDLP
   if (is_distributed_master()) {
     mgpu_engine_->distributed_compute_A_x();
     return;
   }
+#endif
 
   if (!batch_mode_) {
     if constexpr (std::is_same_v<f_t, double>) {
@@ -1229,23 +1235,29 @@ void pdhg_solver_t<i_t, f_t>::compute_next_primal_dual_solution_reflected(
 
   using f_t2 = typename type_2<f_t>::type;
 
+#ifdef CUOPT_ENABLE_DISTRIBUTED_PDLP
   if (is_distributed_master()) { mgpu_engine_->sync_await_shards(stream_view_); }
+#endif
 
   // Compute next primal solution reflected.
 
   if (should_major) {
     graph_all.run(should_major, [&]() {
+#ifdef CUOPT_ENABLE_DISTRIBUTED_PDLP
       // Adds all the shards streams into the graph capture
       if (is_distributed_master()) { mgpu_engine_->graph_capture_fork_to_shards(stream_view_); }
+#endif
 
       compute_At_y();
 
       if (is_distributed_master()) {
+#ifdef CUOPT_ENABLE_DISTRIBUTED_PDLP
         mgpu_engine_->for_each_shard([](auto& shard) {
           auto& sub_pdlp = *shard.sub_pdlp;
           sub_pdlp.pdhg_solver_.primal_reflected_major_projection_transform(
             sub_pdlp.get_primal_step_size());
         });
+#endif
       } else if (!batch_mode_) {
         primal_reflected_major_projection_transform(primal_step_size);
       } else {
@@ -1307,11 +1319,13 @@ void pdhg_solver_t<i_t, f_t>::compute_next_primal_dual_solution_reflected(
       compute_A_x();
 
       if (is_distributed_master()) {
+#ifdef CUOPT_ENABLE_DISTRIBUTED_PDLP
         mgpu_engine_->for_each_shard([](auto& shard) {
           auto& sub_pdlp = *shard.sub_pdlp;
           sub_pdlp.pdhg_solver_.dual_reflected_major_projection_transform(
             sub_pdlp.get_dual_step_size());
         });
+#endif
       } else if (!batch_mode_) {
         dual_reflected_major_projection_transform(dual_step_size);
       } else {
@@ -1338,13 +1352,17 @@ void pdhg_solver_t<i_t, f_t>::compute_next_primal_dual_solution_reflected(
       // Multi-GPU: close the fork by joining every shard stream back into
       // the master stream so cudaStreamEndCapture sees a single graph
       // spanning all streams.
+#ifdef CUOPT_ENABLE_DISTRIBUTED_PDLP
       if (is_distributed_master()) { mgpu_engine_->graph_capture_join_from_shards(stream_view_); }
+#endif
     });
 
   } else {
     graph_all.run(should_major, [&]() {
+#ifdef CUOPT_ENABLE_DISTRIBUTED_PDLP
       // Same reason as above, adds all the shards streams into the graph capture
       if (is_distributed_master()) { mgpu_engine_->graph_capture_fork_to_shards(stream_view_); }
+#endif
 
       // Compute next primal
       compute_At_y();
@@ -1358,11 +1376,13 @@ void pdhg_solver_t<i_t, f_t>::compute_next_primal_dual_solution_reflected(
 #endif
 
       if (is_distributed_master()) {
+#ifdef CUOPT_ENABLE_DISTRIBUTED_PDLP
         mgpu_engine_->for_each_shard([](auto& shard) {
           auto& sub_pdlp = *shard.sub_pdlp;
           sub_pdlp.pdhg_solver_.primal_reflected_projection_transform(
             sub_pdlp.get_primal_step_size());
         });
+#endif
       } else if (!batch_mode_) {
         primal_reflected_projection_transform(primal_step_size);
       } else {
@@ -1425,10 +1445,12 @@ void pdhg_solver_t<i_t, f_t>::compute_next_primal_dual_solution_reflected(
       compute_A_x();
 
       if (is_distributed_master()) {
+#ifdef CUOPT_ENABLE_DISTRIBUTED_PDLP
         mgpu_engine_->for_each_shard([](auto& shard) {
           auto& sub_pdlp = *shard.sub_pdlp;
           sub_pdlp.pdhg_solver_.dual_reflected_projection_transform(sub_pdlp.get_dual_step_size());
         });
+#endif
       } else if (!batch_mode_) {
         dual_reflected_projection_transform(dual_step_size);
       } else {
@@ -1449,12 +1471,16 @@ void pdhg_solver_t<i_t, f_t>::compute_next_primal_dual_solution_reflected(
       print("reflected_dual_", reflected_dual_);
 #endif
 
+#ifdef CUOPT_ENABLE_DISTRIBUTED_PDLP
       if (is_distributed_master()) { mgpu_engine_->graph_capture_join_from_shards(stream_view_); }
+#endif
     });
   }
 
   // sync to master stream after the graph is captured
+#ifdef CUOPT_ENABLE_DISTRIBUTED_PDLP
   if (is_distributed_master()) { mgpu_engine_->sync_await_master(stream_view_); }
+#endif
 }
 
 template <typename i_t, typename f_t>