fix device memory allocation strategy

cpp/src/neighbors/detail/cagra/cagra_merge.cuh

@@ -30,8 +30,6 @@
 #include <cuvs/neighbors/ivf_pq.hpp>
 #include <cuvs/neighbors/refine.hpp>
 
-#include <cuvs/neighbors/nn_descent.hpp>
-
 #include <rmm/resource_ref.hpp>
 
 #include <chrono>
@@ -75,10 +73,6 @@ index<T, IdxT> merge(raft::resources const& handle,
 
   IdxT offset = 0;
 
-  // Allocate the new dataset on device
-  bool dataset_on_device = cuvs::neighbors::nn_descent::has_enough_device_memory(
-    handle, raft::make_extents<std::int64_t>(new_dataset_size, dim), sizeof(IdxT));
-
   auto merge_dataset = [&](T* dst) {
     for (auto index : indices) {
       using ds_idx_type  = decltype(index->data().n_rows());
@@ -99,9 +93,7 @@ index<T, IdxT> merge(raft::resources const& handle,
 
   cagra::index_params output_index_params = params.output_index_params;
 
-  if (dataset_on_device) {
-    RAFT_LOG_DEBUG("cagra merge: using device memory for merged dataset");
-
+  try {
     auto updated_dataset = raft::make_device_matrix<T, std::int64_t>(
       handle, std::int64_t(new_dataset_size), std::int64_t(dim));
 
@@ -118,9 +110,10 @@ index<T, IdxT> merge(raft::resources const& handle,
                                                   std::array<int64_t, 2>{stride, 1});
       merged_index.update_dataset(handle, owning_t{std::move(updated_dataset), out_layout});
     }
+    RAFT_LOG_DEBUG("cagra merge: using device memory for merged dataset");
     return merged_index;
 
-  } else {
+  } catch (std::bad_alloc& e) {
     RAFT_LOG_DEBUG("cagra::merge: using host memory for merged dataset");
 
     auto updated_dataset =

cpp/tests/neighbors/ann_cagra.cuh

@@ -936,10 +936,21 @@ class AnnCagraIndexMergeTest : public ::testing::TestWithParam<AnnCagraInputs> {
  protected:
   void testCagra()
   {
-    // TODO (rhdong): remove when NN Descent index building support InnerProduct. Reference
+    // TODO (tarang-jain): remove when NN Descent index building support InnerProduct. Reference
     // issue: https://github.com/rapidsai/raft/issues/2276
     if (ps.metric == InnerProduct && ps.build_algo == graph_build_algo::NN_DESCENT) GTEST_SKIP();
     if (ps.compression != std::nullopt) GTEST_SKIP();
+    // IVF_PQ and NN_DESCENT graph builds do not support BitwiseHamming
+    if (ps.metric == cuvs::distance::DistanceType::BitwiseHamming &&
+        ((!std::is_same_v<DataT, uint8_t>) ||
+         (ps.build_algo != graph_build_algo::ITERATIVE_CAGRA_SEARCH)))
+      GTEST_SKIP();
+    // If the dataset dimension is small and the dataset size is large, there can be a lot of
+    // dataset vectors that have the same distance to the query, especially in the binary Hamming
+    // distance, making it impossible to make a top-k ground truth.
+    if (ps.metric == cuvs::distance::DistanceType::BitwiseHamming &&
+        (ps.k * ps.dim * 8 / 5 /*(=magic number)*/ < ps.n_rows))
+      GTEST_SKIP();
 
     size_t queries_size = ps.n_queries * ps.k;
     std::vector<IdxT> indices_Cagra(queries_size);