add simplicial_set_embedding to umap

docs/release-notes/0.10.11.md

@@ -3,7 +3,6 @@
 ```{rubric} Features
 ```
 
-
 ```{rubric} Performance
 ```
 
@@ -16,3 +15,4 @@
 * Adds support for `rapids-24.10` {pr}`277` {smaller}`S Dicks`
 * Updates testing to use `rapids-24.10` {pr}`277` {smaller}`S Dicks`
 * Updates `docker` image to use `rapids-24.10` {pr}`277` {smaller}`S Dicks`
+* `tl.UMAP` now uses `simplicial_set_embedding` for `rapids>=24.10` {pr}`278` {smaller}`S Dicks`

src/rapids_singlecell/tools/_umap.py

@@ -2,8 +2,13 @@
 
 from typing import TYPE_CHECKING, Literal
 
+import cuml
+import cupy as cp
+from cuml.manifold.simpl_set import simplicial_set_embedding
 from cuml.manifold.umap import UMAP
 from cuml.manifold.umap_utils import find_ab_params
+from cupyx.scipy import sparse
+from packaging.version import parse as parse_version
 from scanpy._utils import NeighborsView
 from sklearn.utils import check_random_state
 
@@ -149,37 +154,62 @@ def umap(
     )  # 0 is not a valid value for rapids, unlike original umap
 
     n_obs = adata.shape[0]
-    n_neighbors = neigh_params["n_neighbors"]
-    if neigh_params.get("method") == "rapids":
-        knn_dist = neighbors["distances"].data.reshape(n_obs, n_neighbors)
-        knn_indices = neighbors["distances"].indices.reshape(n_obs, n_neighbors)
-        pre_knn = (knn_indices, knn_dist)
+    if parse_version(cuml.__version__) < parse_version("24.10"):
+        # `simplicial_set_embedding` is bugged in cuml<24.10. This is why we use `UMAP` instead.
+        n_neighbors = neigh_params["n_neighbors"]
+        if neigh_params.get("method") == "rapids":
+            knn_dist = neighbors["distances"].data.reshape(n_obs, n_neighbors)
+            knn_indices = neighbors["distances"].indices.reshape(n_obs, n_neighbors)
+            pre_knn = (knn_indices, knn_dist)
+        else:
+            pre_knn = None
+
+        if init_pos == "auto":
+            init_pos = "spectral" if n_obs < 1000000 else "random"
+
+        umap = UMAP(
+            n_neighbors=n_neighbors,
+            n_components=n_components,
+            metric=neigh_params.get("metric", "euclidean"),
+            metric_kwds=neigh_params.get("metric_kwds", None),
+            n_epochs=n_epochs,
+            learning_rate=alpha,
+            init=init_pos,
+            min_dist=min_dist,
+            spread=spread,
+            negative_sample_rate=negative_sample_rate,
+            a=a,
+            b=b,
+            random_state=random_state,
+            output_type="numpy",
+            precomputed_knn=pre_knn,
+        )
+
+        X_umap = umap.fit_transform(X)
     else:
-        pre_knn = None
-
-    if init_pos == "auto":
-        init_pos = "spectral" if n_obs < 1000000 else "random"
-
-    umap = UMAP(
-        n_neighbors=n_neighbors,
-        n_components=n_components,
-        metric=neigh_params.get("metric", "euclidean"),
-        metric_kwds=neigh_params.get("metric_kwds", None),
-        n_epochs=n_epochs,
-        learning_rate=alpha,
-        init=init_pos,
-        min_dist=min_dist,
-        spread=spread,
-        negative_sample_rate=negative_sample_rate,
-        a=a,
-        b=b,
-        random_state=random_state,
-        output_type="numpy",
-        precomputed_knn=pre_knn,
-    )
+        pre_knn = neighbors["connectivities"]
+
+        if init_pos == "auto":
+            init_pos = "spectral" if n_obs < 1000000 else "random"
+
+        X_umap = simplicial_set_embedding(
+            data=cp.array(X),
+            graph=sparse.coo_matrix(pre_knn),
+            n_components=n_components,
+            initial_alpha=alpha,
+            a=a,
+            b=b,
+            negative_sample_rate=negative_sample_rate,
+            n_epochs=n_epochs,
+            init=init_pos,
+            random_state=random_state,
+            metric=neigh_params.get("metric", "euclidean"),
+            metric_kwds=neigh_params.get("metric_kwds", None),
+        )
+        X_umap = cp.asarray(X_umap).get()
 
     key_obsm, key_uns = ("X_umap", "umap") if key_added is None else [key_added] * 2
-    adata.obsm[key_obsm] = umap.fit_transform(X)
+    adata.obsm[key_obsm] = X_umap
 
     adata.uns[key_uns] = {"params": stored_params}
     return adata if copy else None