feat: GPU_HNSW / GPU_HNSW_SQ index using faiss::gpu::GpuIndexHNSW — native FP32/FP16/BF16/INT8

cmake/libs/libdiskann.cmake

@@ -1,6 +1,6 @@
 add_definitions(-DKNOWHERE_WITH_DISKANN)
-find_package(Boost REQUIRED COMPONENTS program_options)
-include_directories(${Boost_INCLUDE_DIR})
+find_package(Boost REQUIRED COMPONENTS program_options CONFIG)
+include_directories(${Boost_INCLUDE_DIRS})
 find_package(aio REQUIRED)
 include_directories(${AIO_INCLUDE})
 find_package(fmt REQUIRED)

cmake/libs/libfaiss.cmake

@@ -593,3 +593,41 @@ if(__PPC64)
                                       knowhere_utils)
   target_compile_definitions(faiss PRIVATE FINTEGER=int)
 endif()
+
+# GPU HNSW CUDA sources — compiled when WITH_CUVS is enabled.
+#
+# This `faiss_gpu_hnsw` OBJECT library is the ONLY place the faiss GPU sources
+# are compiled in the knowhere build:
+#   * The main `faiss` STATIC target above is built from FAISS_SRCS, which globs
+#     only thirdparty/faiss/faiss/{*.cpp,impl,utils,...} — it deliberately does
+#     NOT include thirdparty/faiss/faiss/gpu/*, so none of the files below are
+#     also compiled into `faiss`.
+#   * Upstream faiss ships its own GPU build file
+#     (thirdparty/faiss/faiss/gpu/CMakeLists.txt → the `faiss_gpu_objs` target
+#     over FAISS_GPU_SRC). Knowhere assembles faiss manually via this .cmake and
+#     never `add_subdirectory()`s the vendored faiss tree, so that file (and
+#     `faiss_gpu_objs`) is never part of the knowhere build graph.
+# Therefore the two GPU build paths are mutually exclusive here — there is no
+# duplicate compilation or ODR hazard. Keep it that way: add GPU sources to the
+# list below, NOT by pulling in the vendored gpu/CMakeLists.txt.
+if(WITH_CUVS)
+  set(FAISS_GPU_HNSW_SRCS
+    thirdparty/faiss/faiss/gpu/GpuIndexHNSW.cu
+    thirdparty/faiss/faiss/gpu/GpuIndex.cu
+    thirdparty/faiss/faiss/gpu/GpuResources.cpp
+    thirdparty/faiss/faiss/gpu/StandardGpuResources.cpp
+    thirdparty/faiss/faiss/gpu/impl/GpuHnswTypes.cu
+    thirdparty/faiss/faiss/gpu/impl/IndexUtils.cu
+    thirdparty/faiss/faiss/gpu/utils/DeviceUtils.cu
+    thirdparty/faiss/faiss/gpu/utils/StackDeviceMemory.cpp
+    thirdparty/faiss/faiss/gpu/utils/Timer.cpp
+  )
+  add_library(faiss_gpu_hnsw OBJECT ${FAISS_GPU_HNSW_SRCS})
+  target_include_directories(faiss_gpu_hnsw PRIVATE
+    ${CMAKE_CURRENT_SOURCE_DIR}/thirdparty/faiss
+    ${Boost_INCLUDE_DIRS}
+  )
+  target_compile_definitions(faiss_gpu_hnsw PRIVATE FINTEGER=int)
+  target_link_libraries(faiss_gpu_hnsw PRIVATE CUDA::cudart)
+  target_link_libraries(faiss PUBLIC faiss_gpu_hnsw)
+endif()

include/knowhere/bitsetview.h

@@ -69,6 +69,11 @@ class BitsetView {
         return out_ids_ != nullptr;
     }
 
+    size_t
+    id_offset() const {
+        return id_offset_;
+    }
+
     void
     set_out_ids(const uint32_t* out_ids, size_t num_internal_ids,
                 std::optional<size_t> num_filtered_out_ids = std::nullopt) {

include/knowhere/comp/index_param.h

@@ -53,6 +53,8 @@ constexpr const char* INDEX_GPU_BRUTEFORCE = "GPU_BRUTE_FORCE";
 constexpr const char* INDEX_GPU_IVFFLAT = "GPU_IVF_FLAT";
 constexpr const char* INDEX_GPU_IVFPQ = "GPU_IVF_PQ";
 constexpr const char* INDEX_GPU_CAGRA = "GPU_CAGRA";
+constexpr const char* INDEX_GPU_HNSW = "GPU_HNSW";
+constexpr const char* INDEX_GPU_HNSW_SQ = "GPU_HNSW_SQ";
 
 constexpr const char* INDEX_HNSW = "HNSW";
 constexpr const char* INDEX_HNSW_SQ = "HNSW_SQ";

include/knowhere/index/index_static.h

@@ -19,8 +19,16 @@
 namespace knowhere {
 
 struct Resource {
-    uint64_t memoryCost;  // in bytes
+    uint64_t memoryCost;  // retained host memory after load completes, in bytes
     uint64_t diskCost;    // in bytes
+    // Peak transient host memory required *during* load, in bytes. Defaults to 0
+    // so existing indexes are unaffected: consumers should fall back to their
+    // memoryCost-based heuristic when this is 0. Indexes whose peak load
+    // footprint differs from the retained footprint (e.g. GPU_HNSW frees its CPU
+    // copy after uploading to VRAM, so memoryCost≈0 but the peak covers the
+    // download buffer + deserialized CPU index + decode/graph staging) set this
+    // to the peak so the loader reserves enough host RAM and does not OOM.
+    uint64_t maxMemoryCost = 0;
 };
 
 #define DEFINE_HAS_STATIC_FUNC(func_name)                                               \

include/knowhere/index/index_table.h

@@ -82,6 +82,14 @@ static std::set<std::pair<std::string, VecType>> legal_knowhere_index = {
     {IndexEnum::INDEX_GPU_CAGRA, VecType::VECTOR_FLOAT16},
     {IndexEnum::INDEX_GPU_CAGRA, VecType::VECTOR_INT8},
     {IndexEnum::INDEX_GPU_CAGRA, VecType::VECTOR_BINARY},
+    {IndexEnum::INDEX_GPU_HNSW, VecType::VECTOR_FLOAT},
+    {IndexEnum::INDEX_GPU_HNSW, VecType::VECTOR_FLOAT16},
+    {IndexEnum::INDEX_GPU_HNSW, VecType::VECTOR_BFLOAT16},
+    {IndexEnum::INDEX_GPU_HNSW, VecType::VECTOR_INT8},
+    {IndexEnum::INDEX_GPU_HNSW_SQ, VecType::VECTOR_FLOAT},
+    {IndexEnum::INDEX_GPU_HNSW_SQ, VecType::VECTOR_FLOAT16},
+    {IndexEnum::INDEX_GPU_HNSW_SQ, VecType::VECTOR_BFLOAT16},
+    {IndexEnum::INDEX_GPU_HNSW_SQ, VecType::VECTOR_INT8},
 
     // hnsw
     {IndexEnum::INDEX_HNSW, VecType::VECTOR_FLOAT},