Use HNSW GPU Hierarchy by Default

c/src/neighbors/hnsw.cpp

@@ -122,7 +122,7 @@ extern "C" cuvsError_t cuvsHnswIndexParamsCreate(cuvsHnswIndexParams_t* params)
 {
   return cuvs::core::translate_exceptions([=] {
     *params = new cuvsHnswIndexParams{
-      .hierarchy = cuvsHnswHierarchy::NONE, .ef_construction = 200, .num_threads = 0};
+      .hierarchy = cuvsHnswHierarchy::GPU, .ef_construction = 200, .num_threads = 0};
   });
 }
 

c/tests/neighbors/ann_hnsw_c.cu

@@ -102,6 +102,8 @@ TEST(CagraHnswC, BuildSearch)
   hnsw_index->dtype = index->dtype;
   cuvsHnswIndexParams_t hnsw_params;
   cuvsHnswIndexParamsCreate(&hnsw_params);
+  // Use NONE hierarchy since cuvsCagraSerializeToHnswlib creates a base-layer-only index
+  hnsw_params->hierarchy = NONE;
   cuvsHnswDeserialize(res, hnsw_params, "/tmp/cagra_hnswlib.index", 2, L2Expanded, hnsw_index);
 
   // search index

cpp/include/cuvs/neighbors/hnsw.hpp

@@ -40,7 +40,7 @@ enum class HnswHierarchy {
 
 struct index_params : cuvs::neighbors::index_params {
   /** Hierarchy build type for HNSW index when converting from CAGRA index */
-  HnswHierarchy hierarchy = HnswHierarchy::NONE;
+  HnswHierarchy hierarchy = HnswHierarchy::GPU;
   /** Size of the candidate list during hierarchy construction when hierarchy is `CPU`*/
   int ef_construction = 200;
   /** Number of host threads to use to construct hierarchy when hierarchy is `CPU` or `GPU`.

examples/cpp/src/cagra_hnsw_ace_example.cu

@@ -81,7 +81,8 @@ void cagra_build_search_ace(raft::device_resources const& dev_resources,
   // descriptor For in-memory indices: creates HNSW index in memory
   std::cout << "Converting CAGRA index to HNSW" << std::endl;
   hnsw::index_params hnsw_params;
-  auto hnsw_index = hnsw::from_cagra(dev_resources, hnsw_params, index);
+  hnsw_params.hierarchy = hnsw::HnswHierarchy::GPU;  // Offload hierarchy construction to GPU
+  auto hnsw_index       = hnsw::from_cagra(dev_resources, hnsw_params, index);
 
   // HNSW search requires host matrices
   auto queries_host = raft::make_host_matrix<float, int64_t>(n_queries, queries.extent(1));

java/cuvs-java/src/main/java/com/nvidia/cuvs/HnswIndexParams.java

@@ -27,7 +27,12 @@ public enum CuvsHnswHierarchy {
     /**
      * Full hierarchy is built using the CPU
      */
-    CPU(1);
+    CPU(1),
+
+    /**
+     * Full hierarchy is built using the GPU
+     */
+    GPU(2);
 
     /**
      * The value for the enum choice.
@@ -39,7 +44,16 @@ private CuvsHnswHierarchy(int value) {
     }
   };
 
-  private CuvsHnswHierarchy hierarchy = CuvsHnswHierarchy.NONE;
+  /**
+   * Alias for {@link CuvsHnswHierarchy} for convenience.
+   */
+  public static class HnswHierarchy {
+    public static final CuvsHnswHierarchy NONE = CuvsHnswHierarchy.NONE;
+    public static final CuvsHnswHierarchy CPU = CuvsHnswHierarchy.CPU;
+    public static final CuvsHnswHierarchy GPU = CuvsHnswHierarchy.GPU;
+  }
+
+  private CuvsHnswHierarchy hierarchy = CuvsHnswHierarchy.GPU;
   private int efConstruction = 200;
   private int numThreads = 2;
   private int vectorDimension;
@@ -102,7 +116,7 @@ public String toString() {
    */
   public static class Builder {
 
-    private CuvsHnswHierarchy hierarchy = CuvsHnswHierarchy.NONE;
+    private CuvsHnswHierarchy hierarchy = CuvsHnswHierarchy.GPU;
     private int efConstruction = 200;
     private int numThreads = 2;
     private int vectorDimension;

java/cuvs-java/src/main/java22/com/nvidia/cuvs/internal/HnswIndexImpl.java

@@ -219,6 +219,7 @@ private IndexReference deserialize(InputStream inputStream) throws Throwable {
    */
   private CloseableHandle segmentFromIndexParams(HnswIndexParams params) {
     var hnswParams = createHnswIndexParams();
+    cuvsHnswIndexParams.hierarchy(hnswParams.handle(), params.getHierarchy().value);
     cuvsHnswIndexParams.ef_construction(hnswParams.handle(), params.getEfConstruction());
     cuvsHnswIndexParams.num_threads(hnswParams.handle(), params.getNumThreads());
     return hnswParams;

java/cuvs-java/src/test/java/com/nvidia/cuvs/CagraAceBuildAndSearchIT.java

@@ -10,6 +10,7 @@
 import com.carrotsearch.randomizedtesting.RandomizedRunner;
 import com.nvidia.cuvs.CagraIndexParams.CagraGraphBuildAlgo;
 import com.nvidia.cuvs.CagraIndexParams.CuvsDistanceType;
+import com.nvidia.cuvs.HnswIndexParams.HnswHierarchy;
 import java.nio.file.Files;
 import java.nio.file.Path;
 import java.util.Arrays;
@@ -186,7 +187,7 @@ public void testAceDiskBasedBuild() throws Throwable {
         // Convert CAGRA index to HNSW using fromCagra
         // This automatically handles disk-based indices
         HnswIndexParams hnswIndexParams =
-            new HnswIndexParams.Builder().withVectorDimension(2).build();
+            new HnswIndexParams.Builder().withVectorDimension(2).withHierarchy(HnswHierarchy.GPU).build();
 
         try (var hnswIndexSerialized = HnswIndex.fromCagra(hnswIndexParams, index)) {
           var hnswIndexSerializedPath = buildDir.resolve("hnsw_index.bin");

java/cuvs-java/src/test/java/com/nvidia/cuvs/HnswBuildAndSearchIT.java

@@ -8,6 +8,7 @@
 
 import com.nvidia.cuvs.CagraIndexParams.CagraGraphBuildAlgo;
 import com.nvidia.cuvs.CagraIndexParams.CuvsDistanceType;
+import com.nvidia.cuvs.HnswIndexParams.HnswHierarchy;
 import java.lang.invoke.MethodHandles;
 import java.nio.file.Files;
 import java.nio.file.Path;
@@ -70,8 +71,9 @@ private static void indexAndQueryOnce(
         index.serializeToHNSW(outputStream);
       }
 
+      // Use NONE hierarchy since serializeToHNSW creates a base-layer-only index
       HnswIndexParams hnswIndexParams =
-          new HnswIndexParams.Builder().withVectorDimension(2).build();
+          new HnswIndexParams.Builder().withVectorDimension(2).withHierarchy(HnswHierarchy.NONE).build();
       try (var inputStreamHNSW = Files.newInputStream(hnswIndexPath)) {
         var hnswIndex =
             HnswIndex.newBuilder(resources)

java/cuvs-java/src/test/java/com/nvidia/cuvs/HnswRandomizedIT.java

@@ -9,6 +9,7 @@
 import com.carrotsearch.randomizedtesting.RandomizedRunner;
 import com.nvidia.cuvs.CagraIndexParams.CagraGraphBuildAlgo;
 import com.nvidia.cuvs.CagraIndexParams.CuvsDistanceType;
+import com.nvidia.cuvs.HnswIndexParams.HnswHierarchy;
 import java.lang.invoke.MethodHandles;
 import java.nio.file.Files;
 import java.nio.file.Path;
@@ -129,8 +130,9 @@ private void tmpResultsTopKWithRandomValues(boolean useNativeMemoryDataset) thro
           index.serializeToHNSW(outputStream); // fails here
         }
 
+        // Use NONE hierarchy since serializeToHNSW creates a base-layer-only index
         HnswIndexParams hnswIndexParams =
-            new HnswIndexParams.Builder().withVectorDimension(dimensions).build();
+            new HnswIndexParams.Builder().withVectorDimension(dimensions).withHierarchy(HnswHierarchy.NONE).build();
 
         try (var inputStreamHNSW = Files.newInputStream(hnswIndexPath)) {
           HnswIndex hnswIndex =

python/cuvs/cuvs/neighbors/hnsw/hnsw.pyx

@@ -34,8 +34,8 @@ cdef class IndexParams:
 
     Parameters
     ----------
-    hierarchy : string, default = "none" (optional)
-        The hierarchy of the HNSW index. Valid values are ["none", "cpu"].
+    hierarchy : string, default = "gpu" (optional)
+        The hierarchy of the HNSW index. Valid values are ["none", "cpu", "gpu"].
         - "none": No hierarchy is built.
         - "cpu": Hierarchy is built using CPU.
         - "gpu": Hierarchy is built using GPU.
@@ -61,7 +61,7 @@ cdef class IndexParams:
         check_cuvs(cuvsHnswIndexParamsDestroy(self.params))
 
     def __init__(self, *,
-                 hierarchy="none",
+                 hierarchy="gpu",
                  ef_construction=200,
                  num_threads=0):
         if hierarchy == "none":
@@ -72,7 +72,7 @@ cdef class IndexParams:
             self.params.hierarchy = cuvsHnswHierarchy.GPU
         else:
             raise ValueError("Invalid hierarchy type."
-                             " Valid values are 'none' and 'cpu'.")
+                             " Valid values are 'none', 'cpu', and 'gpu'.")
         self.params.ef_construction = ef_construction
         self.params.num_threads = num_threads