IVF-SQ C++ API

cpp/CMakeLists.txt

@@ -487,6 +487,86 @@ if(NOT BUILD_CPU_ONLY)
     OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/generated_kernels/ivf_flat/post_process"
     KERNEL_LINK_LIBRARIES jit_lto_kernel_usage_requirements
   )
+  set(ivf_sq_ns "cuvs::neighbors::ivf_sq::detail")
+  generate_jit_lto_kernels(
+    jit_lto_files
+    NAME_FORMAT "ivf_sq_scan_capacity_@capacity@"
+    MATRIX_JSON_FILE
+      "${CMAKE_CURRENT_SOURCE_DIR}/src/neighbors/ivf_sq/detail/jit_lto_kernels/scan_matrix.json"
+    KERNEL_INPUT_FILE
+      "${CMAKE_CURRENT_SOURCE_DIR}/src/neighbors/ivf_sq/detail/jit_lto_kernels/scan_kernel.cu.in"
+    FRAGMENT_TAG_FORMAT "${ivf_sq_ns}::fragment_tag_ivf_sq_scan<@capacity@>"
+    FRAGMENT_TAG_HEADER_FILES "<cuvs/detail/jit_lto/ivf_sq/scan_fragments.hpp>"
+                              "<cuvs/detail/jit_lto/common_fragments.hpp>"
+    OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/generated_kernels/ivf_sq/scan"
+    KERNEL_LINK_LIBRARIES jit_lto_kernel_usage_requirements
+  )
+  generate_jit_lto_kernels(
+    jit_lto_files
+    NAME_FORMAT "ivf_sq_setup_invariant_smem_metric_@metric_name@"
+    MATRIX_JSON_FILE
+      "${CMAKE_CURRENT_SOURCE_DIR}/src/neighbors/ivf_sq/detail/jit_lto_kernels/setup_invariant_smem_matrix.json"
+    KERNEL_INPUT_FILE
+      "${CMAKE_CURRENT_SOURCE_DIR}/src/neighbors/ivf_sq/detail/jit_lto_kernels/setup_invariant_smem_kernel.cu.in"
+    FRAGMENT_TAG_FORMAT
+      "${ivf_sq_ns}::fragment_tag_setup_invariant_smem<${ivf_sq_ns}::tag_metric_@metric_name@>"
+    FRAGMENT_TAG_HEADER_FILES "<cuvs/detail/jit_lto/ivf_sq/scan_fragments.hpp>"
+    OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/generated_kernels/ivf_sq/setup_invariant_smem"
+    KERNEL_LINK_LIBRARIES jit_lto_kernel_usage_requirements
+  )
+  generate_jit_lto_kernels(
+    jit_lto_files
+    NAME_FORMAT "ivf_sq_setup_per_probe_smem_metric_@metric_name@"
+    MATRIX_JSON_FILE
+      "${CMAKE_CURRENT_SOURCE_DIR}/src/neighbors/ivf_sq/detail/jit_lto_kernels/setup_per_probe_smem_matrix.json"
+    KERNEL_INPUT_FILE
+      "${CMAKE_CURRENT_SOURCE_DIR}/src/neighbors/ivf_sq/detail/jit_lto_kernels/setup_per_probe_smem_kernel.cu.in"
+    FRAGMENT_TAG_FORMAT
+      "${ivf_sq_ns}::fragment_tag_setup_per_probe_smem<${ivf_sq_ns}::tag_metric_@metric_name@>"
+    FRAGMENT_TAG_HEADER_FILES "<cuvs/detail/jit_lto/ivf_sq/scan_fragments.hpp>"
+    OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/generated_kernels/ivf_sq/setup_per_probe_smem"
+    KERNEL_LINK_LIBRARIES jit_lto_kernel_usage_requirements
+  )
+  generate_jit_lto_kernels(
+    jit_lto_files
+    NAME_FORMAT "ivf_sq_accumulate_distance_metric_@metric_name@"
+    MATRIX_JSON_FILE
+      "${CMAKE_CURRENT_SOURCE_DIR}/src/neighbors/ivf_sq/detail/jit_lto_kernels/accumulate_distance_matrix.json"
+    KERNEL_INPUT_FILE
+      "${CMAKE_CURRENT_SOURCE_DIR}/src/neighbors/ivf_sq/detail/jit_lto_kernels/accumulate_distance_kernel.cu.in"
+    FRAGMENT_TAG_FORMAT
+      "${ivf_sq_ns}::fragment_tag_accumulate_distance<${ivf_sq_ns}::tag_metric_@metric_name@>"
+    FRAGMENT_TAG_HEADER_FILES "<cuvs/detail/jit_lto/ivf_sq/scan_fragments.hpp>"
+    OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/generated_kernels/ivf_sq/accumulate_distance"
+    KERNEL_LINK_LIBRARIES jit_lto_kernel_usage_requirements
+  )
+  generate_jit_lto_kernels(
+    jit_lto_files
+    NAME_FORMAT "ivf_sq_finalize_distance_metric_@metric_name@"
+    MATRIX_JSON_FILE
+      "${CMAKE_CURRENT_SOURCE_DIR}/src/neighbors/ivf_sq/detail/jit_lto_kernels/finalize_distance_matrix.json"
+    KERNEL_INPUT_FILE
+      "${CMAKE_CURRENT_SOURCE_DIR}/src/neighbors/ivf_sq/detail/jit_lto_kernels/finalize_distance_kernel.cu.in"
+    FRAGMENT_TAG_FORMAT
+      "${ivf_sq_ns}::fragment_tag_finalize_distance<${ivf_sq_ns}::tag_metric_@metric_name@>"
+    FRAGMENT_TAG_HEADER_FILES "<cuvs/detail/jit_lto/ivf_sq/scan_fragments.hpp>"
+    OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/generated_kernels/ivf_sq/finalize_distance"
+    KERNEL_LINK_LIBRARIES jit_lto_kernel_usage_requirements
+  )
+  generate_jit_lto_kernels(
+    jit_lto_files
+    NAME_FORMAT "ivf_sq_filter_@filter_name@"
+    MATRIX_JSON_FILE
+      "${CMAKE_CURRENT_SOURCE_DIR}/src/neighbors/ivf_sq/detail/jit_lto_kernels/filter_matrix.json"
+    KERNEL_INPUT_FILE
+      "${CMAKE_CURRENT_SOURCE_DIR}/src/neighbors/ivf_sq/detail/jit_lto_kernels/filter_kernel.cu.in"
+    FRAGMENT_TAG_FORMAT
+      "${ivf_sq_ns}::fragment_tag_ivf_sq_filter<${neighbors_ns}::tag_filter_@filter_name@>"
+    FRAGMENT_TAG_HEADER_FILES "<cuvs/detail/jit_lto/ivf_sq/scan_fragments.hpp>"
+                              "<cuvs/detail/jit_lto/common_fragments.hpp>"
+    OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/generated_kernels/ivf_sq/filter"
+    KERNEL_LINK_LIBRARIES jit_lto_kernel_usage_requirements
+  )
   set(ivf_pq_ns "cuvs::neighbors::ivf_pq::detail")
   generate_jit_lto_kernels(
     jit_lto_files
@@ -897,6 +977,11 @@ if(NOT BUILD_CPU_ONLY)
     src/neighbors/ivf_pq/detail/ivf_pq_process_and_fill_codes.cu
     ${ivf_pq_search_inst_files}
     ${ivf_pq_transform_inst_files}
+    src/neighbors/ivf_sq_index.cpp
+    src/neighbors/ivf_sq/ivf_sq_build_extend_float_uint8_t_int64_t.cu
+    src/neighbors/ivf_sq/ivf_sq_build_extend_half_uint8_t_int64_t.cu
+    src/neighbors/ivf_sq/ivf_sq_search_uint8_t_int64_t.cu
+    src/neighbors/ivf_sq/ivf_sq_serialize_uint8_t.cu
     src/neighbors/knn_merge_parts.cu
     src/neighbors/nn_descent.cu
     ${nn_descent_inst_files}

cpp/bench/ann/CMakeLists.txt

@@ -13,6 +13,7 @@ list(APPEND CMAKE_MODULE_PATH "${CUVS_SOURCE_DIR}")
 option(CUVS_ANN_BENCH_USE_FAISS_GPU_FLAT "Include faiss' brute-force knn algorithm in benchmark" ON)
 option(CUVS_ANN_BENCH_USE_FAISS_GPU_IVF_FLAT "Include faiss' ivf flat algorithm in benchmark" ON)
 option(CUVS_ANN_BENCH_USE_FAISS_GPU_IVF_PQ "Include faiss' ivf pq algorithm in benchmark" ON)
+option(CUVS_ANN_BENCH_USE_FAISS_GPU_IVF_SQ "Include faiss' ivf sq algorithm in benchmark" ON)
 option(CUVS_ANN_BENCH_USE_FAISS_GPU_CAGRA "Include faiss' cagra algorithm in benchmark" ON)
 option(CUVS_ANN_BENCH_USE_FAISS_GPU_CAGRA_HNSW
        "Include faiss' cagra algorithm for build and hnsw for search in benchmark" ON
@@ -22,8 +23,10 @@ option(CUVS_ANN_BENCH_USE_FAISS_CPU_IVF_FLAT "Include faiss' cpu ivf flat algori
        ON
 )
 option(CUVS_ANN_BENCH_USE_FAISS_CPU_IVF_PQ "Include faiss' cpu ivf pq algorithm in benchmark" ON)
+option(CUVS_ANN_BENCH_USE_FAISS_CPU_IVF_SQ "Include faiss' cpu ivf sq algorithm in benchmark" ON)
 option(CUVS_ANN_BENCH_USE_FAISS_CPU_HNSW_FLAT "Include faiss' hnsw algorithm in benchmark" ON)
 option(CUVS_ANN_BENCH_USE_CUVS_IVF_FLAT "Include cuVS ivf flat algorithm in benchmark" ON)
+option(CUVS_ANN_BENCH_USE_CUVS_IVF_SQ "Include cuVS ivf sq algorithm in benchmark" ON)
 option(CUVS_ANN_BENCH_USE_CUVS_IVF_PQ "Include cuVS ivf pq algorithm in benchmark" ON)
 option(CUVS_ANN_BENCH_USE_CUVS_CAGRA "Include cuVS CAGRA in benchmark" ON)
 option(CUVS_ANN_BENCH_USE_CUVS_BRUTE_FORCE "Include cuVS brute force knn in benchmark" ON)
@@ -80,6 +83,7 @@ set(CUVS_USE_FAISS_STATIC ON)
 if(BUILD_CPU_ONLY)
   set(CUVS_FAISS_ENABLE_GPU OFF)
   set(CUVS_ANN_BENCH_USE_CUVS_IVF_FLAT OFF)
+  set(CUVS_ANN_BENCH_USE_CUVS_IVF_SQ OFF)
   set(CUVS_ANN_BENCH_USE_CUVS_IVF_PQ OFF)
   set(CUVS_ANN_BENCH_USE_CUVS_CAGRA OFF)
   set(CUVS_ANN_BENCH_USE_CUVS_BRUTE_FORCE OFF)
@@ -97,6 +101,7 @@ set(CUVS_ANN_BENCH_USE_CUVS OFF)
 if(CUVS_ANN_BENCH_USE_CUVS_IVF_PQ
    OR CUVS_ANN_BENCH_USE_CUVS_BRUTE_FORCE
    OR CUVS_ANN_BENCH_USE_CUVS_IVF_FLAT
+   OR CUVS_ANN_BENCH_USE_CUVS_IVF_SQ
    OR CUVS_ANN_BENCH_USE_CUVS_CAGRA
    OR CUVS_ANN_BENCH_USE_CUVS_CAGRA_HNSWLIB
    OR CUVS_KNN_BENCH_USE_CUVS_BRUTE_FORCE
@@ -244,6 +249,12 @@ if(CUVS_ANN_BENCH_USE_CUVS_IVF_FLAT)
   )
 endif()
 
+if(CUVS_ANN_BENCH_USE_CUVS_IVF_SQ)
+  ConfigureAnnBench(
+    NAME CUVS_IVF_SQ PATH src/cuvs/cuvs_benchmark.cu src/cuvs/cuvs_ivf_sq.cu LINKS cuvs
+  )
+endif()
+
 if(CUVS_ANN_BENCH_USE_CUVS_BRUTE_FORCE)
   ConfigureAnnBench(NAME CUVS_BRUTE_FORCE PATH src/cuvs/cuvs_benchmark.cu LINKS cuvs)
 endif()
@@ -309,6 +320,12 @@ if(CUVS_ANN_BENCH_USE_FAISS_CPU_IVF_PQ)
   )
 endif()
 
+if(CUVS_ANN_BENCH_USE_FAISS_CPU_IVF_SQ)
+  ConfigureAnnBench(
+    NAME FAISS_CPU_IVF_SQ PATH src/faiss/faiss_cpu_benchmark.cpp LINKS ${CUVS_FAISS_TARGETS}
+  )
+endif()
+
 if(CUVS_ANN_BENCH_USE_FAISS_CPU_HNSW_FLAT)
   ConfigureAnnBench(
     NAME FAISS_CPU_HNSW_FLAT PATH src/faiss/faiss_cpu_benchmark.cpp LINKS ${CUVS_FAISS_TARGETS}
@@ -329,6 +346,13 @@ if(CUVS_ANN_BENCH_USE_FAISS_GPU_IVF_PQ AND CUVS_FAISS_ENABLE_GPU)
   )
 endif()
 
+if(CUVS_ANN_BENCH_USE_FAISS_GPU_IVF_SQ AND CUVS_FAISS_ENABLE_GPU)
+  ConfigureAnnBench(
+    NAME FAISS_GPU_IVF_SQ PATH src/faiss/faiss_gpu_benchmark.cu LINKS ${CUVS_FAISS_TARGETS}
+    raft::raft
+  )
+endif()
+
 if(CUVS_ANN_BENCH_USE_FAISS_GPU_FLAT AND CUVS_FAISS_ENABLE_GPU)
   ConfigureAnnBench(
     NAME FAISS_GPU_FLAT PATH src/faiss/faiss_gpu_benchmark.cu LINKS ${CUVS_FAISS_TARGETS}

cpp/bench/ann/src/cuvs/cuvs_ann_bench_param_parser.h

@@ -35,6 +35,11 @@ extern template class cuvs::bench::cuvs_cagra<uint8_t, uint32_t>;
 extern template class cuvs::bench::cuvs_cagra<int8_t, uint32_t>;
 #endif
 
+#ifdef CUVS_ANN_BENCH_USE_CUVS_IVF_SQ
+#include "cuvs_ivf_sq_wrapper.h"
+extern template class cuvs::bench::cuvs_ivf_sq<float>;
+extern template class cuvs::bench::cuvs_ivf_sq<half>;
+#endif
 #ifdef CUVS_ANN_BENCH_USE_CUVS_MG
 #include "cuvs_ivf_flat_wrapper.h"
 #include "cuvs_mg_ivf_flat_wrapper.h"
@@ -86,6 +91,26 @@ void parse_search_param(const nlohmann::json& conf,
 }
 #endif
 
+#ifdef CUVS_ANN_BENCH_USE_CUVS_IVF_SQ
+template <typename T>
+void parse_build_param(const nlohmann::json& conf,
+                       typename cuvs::bench::cuvs_ivf_sq<T>::build_param& param)
+{
+  param.n_lists = conf.at("nlist");
+  if (conf.contains("niter")) { param.kmeans_n_iters = conf.at("niter"); }
+  if (conf.contains("ratio")) {
+    param.kmeans_trainset_fraction = 1.0 / static_cast<double>(conf.at("ratio"));
+  }
+}
+
+template <typename T>
+void parse_search_param(const nlohmann::json& conf,
+                        typename cuvs::bench::cuvs_ivf_sq<T>::search_param& param)
+{
+  param.ivf_sq_params.n_probes = conf.at("nprobe");
+}
+#endif
+
 #if defined(CUVS_ANN_BENCH_USE_CUVS_IVF_PQ) || defined(CUVS_ANN_BENCH_USE_CUVS_CAGRA) ||   \
   defined(CUVS_ANN_BENCH_USE_CUVS_CAGRA_HNSWLIB) || defined(CUVS_ANN_BENCH_USE_CUVS_MG) || \
   defined(CUVS_ANN_BENCH_USE_CUVS_CAGRA_DISKANN)

cpp/bench/ann/src/cuvs/cuvs_benchmark.cu

@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2023-2025, NVIDIA CORPORATION.
+ * SPDX-FileCopyrightText: Copyright (c) 2023-2026, NVIDIA CORPORATION.
  * SPDX-License-Identifier: Apache-2.0
  */
 
@@ -84,6 +84,15 @@ auto create_algo(const std::string& algo_name,
     }
   }
 #endif
+#ifdef CUVS_ANN_BENCH_USE_CUVS_IVF_SQ
+  if constexpr (std::is_same_v<T, float> || std::is_same_v<T, half>) {
+    if (algo_name == "cuvs_ivf_sq") {
+      typename cuvs::bench::cuvs_ivf_sq<T>::build_param param;
+      parse_build_param<T>(conf, param);
+      a = std::make_unique<cuvs::bench::cuvs_ivf_sq<T>>(metric, dim, param);
+    }
+  }
+#endif
 #ifdef CUVS_ANN_BENCH_USE_CUVS_IVF_PQ
   if (algo_name == "raft_ivf_pq" || algo_name == "cuvs_ivf_pq") {
     typename cuvs::bench::cuvs_ivf_pq<T, int64_t>::build_param param;
@@ -151,6 +160,15 @@ auto create_search_param(const std::string& algo_name, const nlohmann::json& con
     }
   }
 #endif
+#ifdef CUVS_ANN_BENCH_USE_CUVS_IVF_SQ
+  if constexpr (std::is_same_v<T, float> || std::is_same_v<T, half>) {
+    if (algo_name == "cuvs_ivf_sq") {
+      auto param = std::make_unique<typename cuvs::bench::cuvs_ivf_sq<T>::search_param>();
+      parse_search_param<T>(conf, *param);
+      return param;
+    }
+  }
+#endif
 #ifdef CUVS_ANN_BENCH_USE_CUVS_IVF_PQ
   if (algo_name == "raft_ivf_pq" || algo_name == "cuvs_ivf_pq") {
     auto param = std::make_unique<typename cuvs::bench::cuvs_ivf_pq<T, int64_t>::search_param>();

cpp/bench/ann/src/cuvs/cuvs_ivf_sq.cu

@@ -0,0 +1,10 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+#include "cuvs_ivf_sq_wrapper.h"
+
+namespace cuvs::bench {
+template class cuvs_ivf_sq<float>;
+template class cuvs_ivf_sq<half>;
+}  // namespace cuvs::bench