Merge branch 'main' into test-uv-docs

docs/conf.py

@@ -62,7 +62,7 @@
 
 autosummary_generate = True
 autodoc_member_order = "bysource"
-autodoc_mock_imports = ["cudf", "cuml", "cugraph", "cupy", "cupyx", "pylibraft"]
+autodoc_mock_imports = ["cudf", "cuml", "cugraph", "cupy", "cupyx", "pylibraft", "cuvs"]
 default_role = "literal"
 napoleon_google_docstring = False
 napoleon_numpy_docstring = True

docs/release-notes/0.10.12.md

@@ -2,6 +2,7 @@
 
 ```{rubric} Features
 ```
+* use `cuvs` over `raft` for `pp.neighbors` for `rapids>=24.12`{pr}`304` {smaller}`S Dicks`
 ```{rubric} Performance
 ```
 ```{rubric} Bug fixes

src/rapids_singlecell/preprocessing/_neighbors.py

@@ -6,8 +6,10 @@
 
 import cupy as cp
 import numpy as np
+import pylibraft
 from cuml.manifold.simpl_set import fuzzy_simplicial_set
 from cupyx.scipy import sparse as cp_sparse
+from packaging.version import parse as parse_version
 from pylibraft.common import DeviceResources
 from scipy import sparse as sc_sparse
 
@@ -59,6 +61,10 @@
 _Metrics = _MetricsDense | _MetricsSparse
 
 
+def _cuvs_switch():
+    return parse_version(pylibraft.__version__) > parse_version("24.10")
+
+
 def _brute_knn(
     X: cp_sparse.spmatrix | cp.ndarray,
     Y: cp_sparse.spmatrix | cp.ndarray,
@@ -83,81 +89,114 @@ def _brute_knn(
 def _cagra_knn(
     X: cp.ndarray, Y: cp.ndarray, k: int, metric: _Metrics, metric_kwds: Mapping
 ) -> tuple[cp.ndarray, cp.ndarray]:
-    try:
-        from pylibraft.neighbors import cagra
-    except ImportError:
-        raise ImportError(
-            "The 'cagra' module is not available in your current RAFT installation. "
-            "Please update RAFT to a version that supports 'cagra'."
-        )
+    if not _cuvs_switch():
+        try:
+            from pylibraft.neighbors import cagra
+        except ImportError:
+            raise ImportError(
+                "The 'cagra' module is not available in your current RAFT installation. "
+                "Please update RAFT to a version that supports 'cagra'."
+            )
+        resources = DeviceResources()
+        build_kwargs = {"handle": resources}
+        search_kwargs = {"handle": resources}
+    else:
+        from cuvs.neighbors import cagra
+
+        resources = None
+        build_kwargs = {}
+        search_kwargs = {}
 
-    handle = DeviceResources()
     build_params = cagra.IndexParams(metric="sqeuclidean", build_algo="nn_descent")
-    index = cagra.build(build_params, X, handle=handle)
+    index = cagra.build(build_params, X, **build_kwargs)
 
     n_samples = Y.shape[0]
     all_neighbors = cp.zeros((n_samples, k), dtype=cp.int32)
     all_distances = cp.zeros((n_samples, k), dtype=cp.float32)
 
     batchsize = 65000
-    n_batches = math.ceil(Y.shape[0] / batchsize)
+    n_batches = math.ceil(n_samples / batchsize)
     for batch in range(n_batches):
         start_idx = batch * batchsize
-        stop_idx = min(batch * batchsize + batchsize, Y.shape[0])
+        stop_idx = min((batch + 1) * batchsize, n_samples)
         batch_Y = Y[start_idx:stop_idx, :]
+
         search_params = cagra.SearchParams()
         distances, neighbors = cagra.search(
-            search_params, index, batch_Y, k, handle=handle
+            search_params, index, batch_Y, k, **search_kwargs
         )
         all_neighbors[start_idx:stop_idx, :] = cp.asarray(neighbors)
         all_distances[start_idx:stop_idx, :] = cp.asarray(distances)
-    handle.sync()
+
+    if resources is not None:
+        resources.sync()
+
     if metric == "euclidean":
         all_distances = cp.sqrt(all_distances)
+
     return all_neighbors, all_distances
 
 
 def _ivf_flat_knn(
     X: cp.ndarray, Y: cp.ndarray, k: int, metric: _Metrics, metric_kwds: Mapping
 ) -> tuple[cp.ndarray, cp.ndarray]:
-    from pylibraft.neighbors import ivf_flat
+    if not _cuvs_switch():
+        from pylibraft.neighbors import ivf_flat
 
-    handle = DeviceResources()
-    if X.shape[0] < 2048:
-        n_lists = X.shape[0] // 2
+        resources = DeviceResources()
+        build_kwargs = {"handle": resources}  # pylibraft uses 'handle'
+        search_kwargs = {"handle": resources}
     else:
-        n_lists = 1024
-    index_params = ivf_flat.IndexParams(n_lists=n_lists, metric=metric)
-    index = ivf_flat.build(index_params, X, handle=handle)
+        from cuvs.neighbors import ivf_flat
 
+        resources = None
+        build_kwargs = {}  # cuvs does not need handle/resources
+        search_kwargs = {}
+
+    n_lists = int(math.sqrt(X.shape[0]))
+    index_params = ivf_flat.IndexParams(n_lists=n_lists, metric=metric)
+    index = ivf_flat.build(index_params, X, **build_kwargs)
     distances, neighbors = ivf_flat.search(
-        ivf_flat.SearchParams(), index, Y, k, handle=handle
+        ivf_flat.SearchParams(), index, Y, k, **search_kwargs
     )
+
+    if resources is not None:
+        resources.sync()
+
     distances = cp.asarray(distances)
     neighbors = cp.asarray(neighbors)
-    handle.sync()
+
     return neighbors, distances
 
 
 def _ivf_pq_knn(
     X: cp.ndarray, Y: cp.ndarray, k: int, metric: _Metrics, metric_kwds: Mapping
 ) -> tuple[cp.ndarray, cp.ndarray]:
-    from pylibraft.neighbors import ivf_pq
+    if not _cuvs_switch():
+        from pylibraft.neighbors import ivf_pq
 
-    handle = DeviceResources()
-    if X.shape[0] < 2048:
-        n_lists = X.shape[0] // 2
+        resources = DeviceResources()
+        build_kwargs = {"handle": resources}
+        search_kwargs = {"handle": resources}
     else:
-        n_lists = 1024
-    index_params = ivf_pq.IndexParams(n_lists=n_lists, metric=metric)
-    index = ivf_pq.build(index_params, X, handle=handle)
+        from cuvs.neighbors import ivf_pq
+
+        resources = None
+        build_kwargs = {}
+        search_kwargs = {}
 
+    n_lists = int(math.sqrt(X.shape[0]))
+    index_params = ivf_pq.IndexParams(n_lists=n_lists, metric=metric)
+    index = ivf_pq.build(index_params, X, **build_kwargs)
     distances, neighbors = ivf_pq.search(
-        ivf_pq.SearchParams(), index, Y, k, handle=handle
+        ivf_pq.SearchParams(), index, Y, k, **search_kwargs
     )
+    if resources is not None:
+        resources.sync()
+
     distances = cp.asarray(distances)
     neighbors = cp.asarray(neighbors)
-    handle.sync()
+
     return neighbors, distances