diff --git a/pyproject.toml b/pyproject.toml
index a6af021886..4cf55602f2 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -68,6 +68,7 @@ dependencies = [
     "pynndescent>=0.5.13",
     "scikit-learn>=1.6",
     "scipy>=1.13",
+    "scverse-backends>=0.0.2,<0.1",
     "scverse-misc[settings]>=0.0.5",
     "seaborn>=0.13.2",
     "session-info2",
diff --git a/src/scanpy/_backends.py b/src/scanpy/_backends.py
new file mode 100644
index 0000000000..f9ac431056
--- /dev/null
+++ b/src/scanpy/_backends.py
@@ -0,0 +1,42 @@
+"""Backend dispatch integration for optional Scanpy accelerators."""
+
+from __future__ import annotations
+
+from scverse_backends import BackendDispatcher
+
+dispatcher = BackendDispatcher(
+    entrypoint_group="scanpy.backends",
+    host_name="scanpy",
+    trusted_backends={
+        "rapids_singlecell": {
+            "aliases": ["cuda", "rapids", "rapids-singlecell"],
+            "distributions": [
+                "rapids-singlecell",
+                "rapids-singlecell-cu12",
+                "rapids-singlecell-cu13",
+            ],
+            "package": "rapids-singlecell",
+        },
+    },
+    reserved_backends={
+        "gpu": (
+            "Use 'cuda' for the RAPIDS backend. The generic 'gpu' selector is "
+            "reserved so future GPU backends can coexist without ambiguity."
+        ),
+    },
+)
+
+backend_dispatch = dispatcher.backend_dispatch
+settings = dispatcher.settings
+get_backend = dispatcher.get_backend
+available_backend_names = dispatcher.available_backend_names
+discover = dispatcher.discover
+
+__all__ = [
+    "available_backend_names",
+    "backend_dispatch",
+    "discover",
+    "dispatcher",
+    "get_backend",
+    "settings",
+]
diff --git a/src/scanpy/_settings/__init__.py b/src/scanpy/_settings/__init__.py
index 1fff61f409..0e5a1734e5 100644
--- a/src/scanpy/_settings/__init__.py
+++ b/src/scanpy/_settings/__init__.py
@@ -173,7 +173,39 @@ def model_post_init(self, context: object) -> None:
 
     @computed_field
     @property
-    def logpath(self) -> Path | None:
+    def backend(cls) -> str:
+        """Active computational backend (default ``'cpu'``)."""
+        from .._backends import settings as backend_settings
+
+        return backend_settings.backend
+
+    @backend.setter
+    @_type_check_arg2(str)
+    def backend(cls, backend: str) -> None:
+        from .._backends import settings as backend_settings
+
+        backend_settings.backend = backend
+
+    def use_backend(cls, backend: str):
+        """Temporarily set the active computational backend."""
+        from .._backends import settings as backend_settings
+
+        return backend_settings.use_backend(backend)
+
+    def available_backends(cls) -> list[str]:
+        """Return canonical names of installed computational backends."""
+        from .._backends import settings as backend_settings
+
+        return backend_settings.available_backends()
+
+    def get_backend(cls, name: str):
+        """Look up an installed computational backend by name or alias."""
+        from .._backends import settings as backend_settings
+
+        return backend_settings.get_backend(name)
+
+    @property
+    def logpath(cls) -> Path | None:
         """The file path `logfile` was set to."""
         if self.logfile is _default_logfile():
             return None
diff --git a/src/scanpy/get/_aggregated.py b/src/scanpy/get/_aggregated.py
index fe5140171b..0390d18b3f 100644
--- a/src/scanpy/get/_aggregated.py
+++ b/src/scanpy/get/_aggregated.py
@@ -10,6 +10,7 @@
 from scipy import sparse
 from sklearn.utils.sparsefuncs import csc_median_axis_0
 
+from scanpy._backends import backend_dispatch
 from scanpy._compat import CSBase, CSRBase, DaskArray
 
 from .._utils import _resolve_axis, get_literal_vals
@@ -197,6 +198,7 @@ def _power(x: Array, power: float) -> Array:
     return x**power if isinstance(x, np.ndarray) else x.power(power)
 
 
+@backend_dispatch
 def aggregate(  # noqa: PLR0912
     adata: AnnData,
     by: str | Collection[str],
diff --git a/src/scanpy/neighbors/__init__.py b/src/scanpy/neighbors/__init__.py
index 3622c02576..faadb5c966 100644
--- a/src/scanpy/neighbors/__init__.py
+++ b/src/scanpy/neighbors/__init__.py
@@ -15,6 +15,7 @@
 
 from .. import _utils
 from .. import logging as logg
+from .._backends import backend_dispatch
 from .._compat import CSBase, CSRBase, SpBase, pkg_version, warn
 from .._docs import doc_rng
 from .._settings import settings
@@ -82,8 +83,9 @@ class NeighborsParams(TypedDict):  # noqa: D101
     n_pcs: NotRequired[int]
 
 
-@_doc_params(n_pcs=doc_n_pcs, use_rep=doc_use_rep, rng=doc_rng)
 @_accepts_legacy_random_state(_DEFAULT_SEED := 0)
+@backend_dispatch
+@_doc_params(n_pcs=doc_n_pcs, use_rep=doc_use_rep, rng=doc_rng)
 def neighbors(  # noqa: PLR0913
     adata: AnnData,
     n_neighbors: int = 15,
diff --git a/src/scanpy/preprocessing/_harmony/__init__.py b/src/scanpy/preprocessing/_harmony/__init__.py
index 98b0c6f5fe..98d8e7dbdc 100644
--- a/src/scanpy/preprocessing/_harmony/__init__.py
+++ b/src/scanpy/preprocessing/_harmony/__init__.py
@@ -4,6 +4,7 @@
 
 import numpy as np
 
+from ..._backends import backend_dispatch
 from ..._compat import warn
 
 if TYPE_CHECKING:
@@ -16,6 +17,7 @@
     from ..._utils.random import RNGLike, SeedLike
 
 
+@backend_dispatch
 def harmony_integrate(  # noqa: PLR0913
     adata: AnnData,
     key: str | Sequence[str],
diff --git a/src/scanpy/preprocessing/_highly_variable_genes.py b/src/scanpy/preprocessing/_highly_variable_genes.py
index d5f3d2cc79..1ff712299e 100644
--- a/src/scanpy/preprocessing/_highly_variable_genes.py
+++ b/src/scanpy/preprocessing/_highly_variable_genes.py
@@ -13,6 +13,7 @@
 from fast_array_utils import stats
 
 from .. import logging as logg
+from .._backends import backend_dispatch
 from .._compat import CSBase, CSRBase, DaskArray, warn
 from .._settings import Default, Verbosity, settings
 from .._utils import (
@@ -591,6 +592,7 @@ def _highly_variable_genes_batched(
     return df
 
 
+@backend_dispatch
 def highly_variable_genes(  # noqa: PLR0913
     adata: AnnData,
     *,
diff --git a/src/scanpy/preprocessing/_normalization.py b/src/scanpy/preprocessing/_normalization.py
index 28d7c20293..7b92f8ac71 100644
--- a/src/scanpy/preprocessing/_normalization.py
+++ b/src/scanpy/preprocessing/_normalization.py
@@ -9,6 +9,7 @@
 from fast_array_utils.numba import njit
 
 from .. import logging as logg
+from .._backends import backend_dispatch
 from .._compat import CSBase, CSCBase, CSRBase, DaskArray, warn
 from .._utils import axis_mul_or_truediv, dematrix, view_to_actual
 from ..get import _get_obs_rep, _set_obs_rep
@@ -124,6 +125,7 @@ def _normalize_total_helper(
     return x, counts_per_cell, gene_subset
 
 
+@backend_dispatch
 def normalize_total(  # noqa: PLR0912
     adata: AnnData,
     *,
diff --git a/src/scanpy/preprocessing/_pca/__init__.py b/src/scanpy/preprocessing/_pca/__init__.py
index 4cbcf4c7e6..9ed296a1f8 100644
--- a/src/scanpy/preprocessing/_pca/__init__.py
+++ b/src/scanpy/preprocessing/_pca/__init__.py
@@ -7,6 +7,7 @@
 from packaging.version import Version
 
 from ... import logging as logg
+from ..._backends import backend_dispatch
 from ..._compat import CSBase, DaskArray, pkg_version, warn
 from ..._docs import doc_rng
 from ..._settings import Default, settings
@@ -51,8 +52,9 @@
 type SvdSolver = SvdSolvDaskML | SvdSolvSkearn | SvdSolvPCACustom
 
 
-@_doc_params(mask_var=doc_mask_var, rng=doc_rng)
 @_accepts_legacy_random_state(0)
+@backend_dispatch
+@_doc_params(mask_var=doc_mask_var, rng=doc_rng)
 def pca(  # noqa: PLR0912, PLR0913, PLR0915
     data: AnnData | np.ndarray | CSBase,
     n_comps: int | None = None,
diff --git a/src/scanpy/preprocessing/_qc.py b/src/scanpy/preprocessing/_qc.py
index 04340ec7e9..5710322ec3 100644
--- a/src/scanpy/preprocessing/_qc.py
+++ b/src/scanpy/preprocessing/_qc.py
@@ -13,6 +13,7 @@
 from scanpy.get import _get_obs_rep
 from scanpy.preprocessing._distributed import materialize_as_ndarray
 
+from .._backends import backend_dispatch
 from .._compat import CSBase, CSRBase, DaskArray, warn
 from .._utils import _doc_params, axis_nnz
 from ._docs import (
@@ -204,6 +205,7 @@ def describe_var(
     doc_obs_qc_returns=doc_obs_qc_returns,
     doc_var_qc_returns=doc_var_qc_returns,
 )
+@backend_dispatch
 def calculate_qc_metrics(
     adata: AnnData,
     *,
diff --git a/src/scanpy/preprocessing/_scale.py b/src/scanpy/preprocessing/_scale.py
index 1a1a047544..451b3376a1 100644
--- a/src/scanpy/preprocessing/_scale.py
+++ b/src/scanpy/preprocessing/_scale.py
@@ -11,6 +11,7 @@
 from fast_array_utils.stats import mean_var
 
 from .. import logging as logg
+from .._backends import backend_dispatch
 from .._compat import CSBase, CSCBase, CSRBase, DaskArray, warn
 from .._settings import Default, settings
 from .._utils import (
@@ -68,6 +69,7 @@ def clip_array(
     return x
 
 
+@backend_dispatch
 @singledispatch
 def scale[A: _Array](
     data: AnnData | A,
diff --git a/src/scanpy/preprocessing/_scrublet/__init__.py b/src/scanpy/preprocessing/_scrublet/__init__.py
index 6aae43c608..6036927e45 100644
--- a/src/scanpy/preprocessing/_scrublet/__init__.py
+++ b/src/scanpy/preprocessing/_scrublet/__init__.py
@@ -10,6 +10,7 @@
 
 from ... import logging as logg
 from ... import preprocessing as pp
+from ..._backends import backend_dispatch
 from ..._docs import doc_rng
 from ..._utils import _doc_params
 from ..._utils.random import _accepts_legacy_random_state, _LegacyRng
@@ -23,6 +24,7 @@
 
 
 @_accepts_legacy_random_state(0)
+@backend_dispatch
 @_doc_params(rng=doc_rng)
 def scrublet(  # noqa: PLR0913
     adata: AnnData,
@@ -497,6 +499,7 @@ def _scrublet_call_doublets(  # noqa: PLR0913
 
 
 @_accepts_legacy_random_state(0)
+@backend_dispatch
 def scrublet_simulate_doublets(
     adata: AnnData,
     *,
diff --git a/src/scanpy/preprocessing/_simple.py b/src/scanpy/preprocessing/_simple.py
index fb325dec35..a46cd942cd 100644
--- a/src/scanpy/preprocessing/_simple.py
+++ b/src/scanpy/preprocessing/_simple.py
@@ -23,6 +23,7 @@
 from sklearn.utils import check_array
 
 from .. import logging as logg
+from .._backends import backend_dispatch
 from .._compat import CSBase, CSRBase, DaskArray
 from .._docs import doc_rng
 from .._settings import settings
@@ -50,6 +51,7 @@
     from .._utils.random import RNGLike, SeedLike
 
 
+@backend_dispatch
 def filter_cells(
     data: AnnData | CSBase | np.ndarray | DaskArray,
     *,
@@ -196,6 +198,7 @@ def filter_cells(
     return cell_subset, number_per_cell
 
 
+@backend_dispatch
 def filter_genes(
     data: AnnData | CSBase | np.ndarray | DaskArray,
     *,
@@ -306,6 +309,7 @@ def filter_genes(
     return gene_subset, number_per_gene
 
 
+@backend_dispatch
 @singledispatch
 def log1p(
     data: AnnData | np.ndarray | CSBase,
@@ -507,6 +511,7 @@ def numpy_regress_out(
     return data
 
 
+@backend_dispatch
 def regress_out(
     adata: AnnData,
     keys: str | Sequence[str],
diff --git a/src/scanpy/testing/__init__.py b/src/scanpy/testing/__init__.py
new file mode 100644
index 0000000000..42a4700282
--- /dev/null
+++ b/src/scanpy/testing/__init__.py
@@ -0,0 +1,7 @@
+"""Testing helpers for Scanpy extensions."""
+
+from __future__ import annotations
+
+from testing.scanpy._backends import validate_backend
+
+__all__ = ["validate_backend"]
diff --git a/src/scanpy/tools/_diffmap.py b/src/scanpy/tools/_diffmap.py
index 90f2976837..0817448396 100644
--- a/src/scanpy/tools/_diffmap.py
+++ b/src/scanpy/tools/_diffmap.py
@@ -4,6 +4,7 @@
 
 import numpy as np
 
+from .._backends import backend_dispatch
 from .._docs import doc_rng
 from .._utils import _doc_params
 from .._utils.random import _accepts_legacy_random_state
@@ -15,8 +16,9 @@
     from .._utils.random import RNGLike, SeedLike
 
 
-@_doc_params(rng=doc_rng)
 @_accepts_legacy_random_state(0)
+@backend_dispatch
+@_doc_params(rng=doc_rng)
 def diffmap(
     adata: AnnData,
     n_comps: int = 15,
diff --git a/src/scanpy/tools/_draw_graph.py b/src/scanpy/tools/_draw_graph.py
index bdbb37d0be..7036309f62 100644
--- a/src/scanpy/tools/_draw_graph.py
+++ b/src/scanpy/tools/_draw_graph.py
@@ -7,6 +7,7 @@
 
 from .. import _utils
 from .. import logging as logg
+from .._backends import backend_dispatch
 from .._docs import doc_rng
 from .._utils import _choose_graph, _doc_params, get_literal_vals
 from .._utils.random import (
@@ -31,8 +32,9 @@
 type _Layout = Literal["fr", "drl", "kk", "grid_fr", "lgl", "rt", "rt_circular", "fa"]
 
 
-@_doc_params(rng=doc_rng)
 @_accepts_legacy_random_state(0)
+@backend_dispatch
+@_doc_params(rng=doc_rng)
 def draw_graph(  # noqa: PLR0913
     adata: AnnData,
     layout: _Layout = "fa",
diff --git a/src/scanpy/tools/_embedding_density.py b/src/scanpy/tools/_embedding_density.py
index 37e2bcf606..b0f1d09d14 100644
--- a/src/scanpy/tools/_embedding_density.py
+++ b/src/scanpy/tools/_embedding_density.py
@@ -7,6 +7,7 @@
 import numpy as np
 
 from .. import logging as logg
+from .._backends import backend_dispatch
 from .._utils import sanitize_anndata
 
 if TYPE_CHECKING:
@@ -32,6 +33,7 @@ def _calc_density(x: np.ndarray, y: np.ndarray):
     return scaled_z
 
 
+@backend_dispatch
 def embedding_density(  # noqa: PLR0912
     adata: AnnData,
     basis: str = "umap",
diff --git a/src/scanpy/tools/_leiden.py b/src/scanpy/tools/_leiden.py
index 2f789acbaf..09ea969f21 100644
--- a/src/scanpy/tools/_leiden.py
+++ b/src/scanpy/tools/_leiden.py
@@ -8,6 +8,7 @@
 
 from .. import _utils
 from .. import logging as logg
+from .._backends import backend_dispatch
 from .._compat import warn
 from .._docs import doc_rng
 from .._settings import Default
@@ -43,6 +44,8 @@
             )
 
 
+@_accepts_legacy_random_state(0)
+@backend_dispatch
 @_doc_params(
     doc_adata=doc_adata,
     restrict_to=doc_restrict_to,
@@ -51,7 +54,6 @@
     obsp=doc_obsp,
     rng=doc_rng,
 )
-@_accepts_legacy_random_state(0)
 def leiden(  # noqa: PLR0913
     adata: AnnData,
     resolution: float = 1,
diff --git a/src/scanpy/tools/_louvain.py b/src/scanpy/tools/_louvain.py
index 19fb50bd47..f7138f15cf 100644
--- a/src/scanpy/tools/_louvain.py
+++ b/src/scanpy/tools/_louvain.py
@@ -11,7 +11,8 @@
 
 from .. import _utils
 from .. import logging as logg
-from .._compat import pkg_version
+from .._backends import backend_dispatch
+from .._compat import deprecated, pkg_version
 from .._utils import _choose_graph, _doc_params
 from ._docs import (
     doc_adata,
@@ -40,6 +41,7 @@
             )
 
 
+@backend_dispatch
 @deprecated(Deprecation("1.12.0", "Use :func:`scanpy.tl.leiden` instead."))
 @_doc_params(
     doc_adata=doc_adata,
diff --git a/src/scanpy/tools/_rank_genes_groups.py b/src/scanpy/tools/_rank_genes_groups.py
index 0e6bb73386..85240dc186 100644
--- a/src/scanpy/tools/_rank_genes_groups.py
+++ b/src/scanpy/tools/_rank_genes_groups.py
@@ -13,6 +13,7 @@
 
 from .. import _utils
 from .. import logging as logg
+from .._backends import backend_dispatch
 from .._compat import CSBase
 from .._settings import Default
 from .._settings.presets import DETest
@@ -493,6 +494,7 @@ def compute_statistics(  # noqa: PLR0912
             self.stats.index = self.var_names
 
 
+@backend_dispatch
 def rank_genes_groups(  # noqa: PLR0912, PLR0913, PLR0915
     adata: AnnData,
     groupby: str,
diff --git a/src/scanpy/tools/_score_genes.py b/src/scanpy/tools/_score_genes.py
index 6ba7f51a4c..69ba104bd5 100644
--- a/src/scanpy/tools/_score_genes.py
+++ b/src/scanpy/tools/_score_genes.py
@@ -8,6 +8,7 @@
 import pandas as pd
 
 from .. import logging as logg
+from .._backends import backend_dispatch
 from .._compat import CSBase
 from .._docs import doc_rng
 from .._settings import Default, settings
@@ -53,8 +54,9 @@ def _sparse_nanmean(x: CSBase, /, axis: Literal[0, 1]) -> NDArray[np.float64]:
     return m
 
 
-@_doc_params(rng=doc_rng)
 @_accepts_legacy_random_state(0)
+@backend_dispatch
+@_doc_params(rng=doc_rng)
 def score_genes(  # noqa: PLR0913
     adata: AnnData,
     gene_list: Sequence[str] | pd.Index[str],
@@ -267,6 +269,7 @@ def _nan_means(
     return np.nanmean(x, axis=axis, dtype=dtype)
 
 
+@backend_dispatch
 def score_genes_cell_cycle(
     adata: AnnData,
     *,
diff --git a/src/scanpy/tools/_tsne.py b/src/scanpy/tools/_tsne.py
index 54fb28f436..c8d93d04b5 100644
--- a/src/scanpy/tools/_tsne.py
+++ b/src/scanpy/tools/_tsne.py
@@ -3,6 +3,7 @@
 from typing import TYPE_CHECKING
 
 from .. import logging as logg
+from .._backends import backend_dispatch
 from .._compat import warn
 from .._docs import doc_rng
 from .._settings import settings
@@ -18,6 +19,7 @@
 
 
 @_accepts_legacy_random_state(0)
+@backend_dispatch
 @_doc_params(doc_n_pcs=doc_n_pcs, use_rep=doc_use_rep, rng=doc_rng)
 def tsne(  # noqa: PLR0913
     adata: AnnData,
diff --git a/src/scanpy/tools/_umap.py b/src/scanpy/tools/_umap.py
index 316c6c8aed..9bc95bce90 100644
--- a/src/scanpy/tools/_umap.py
+++ b/src/scanpy/tools/_umap.py
@@ -6,6 +6,7 @@
 from sklearn.utils import check_array
 
 from .. import logging as logg
+from .._backends import backend_dispatch
 from .._docs import doc_rng
 from .._settings import settings
 from .._utils import NeighborsView, _doc_params
@@ -28,6 +29,7 @@
 
 
 @_accepts_legacy_random_state(0)
+@backend_dispatch
 @_doc_params(rng=doc_rng)
 def umap(  # noqa: PLR0913
     adata: AnnData,
diff --git a/src/testing/scanpy/__init__.py b/src/testing/scanpy/__init__.py
index 08571d88a9..3995ccc90e 100644
--- a/src/testing/scanpy/__init__.py
+++ b/src/testing/scanpy/__init__.py
@@ -1,3 +1,18 @@
 """Scanpy testing utilities."""
 
-# This file is empty until we design its public API.
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from ._backends import validate_backend
+
+__all__ = ["validate_backend"]
+
+
+def __getattr__(name: str):
+    if name == "validate_backend":
+        from ._backends import validate_backend
+
+        return validate_backend
+    raise AttributeError(name)
diff --git a/src/testing/scanpy/_backends.py b/src/testing/scanpy/_backends.py
new file mode 100644
index 0000000000..87b9438c6b
--- /dev/null
+++ b/src/testing/scanpy/_backends.py
@@ -0,0 +1,102 @@
+"""Backend conformance helpers for Scanpy-compatible accelerators."""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import numpy as np
+from anndata import AnnData
+from scverse_backends.testing import run_conformance
+
+import scanpy as sc
+from scanpy._backends import get_backend
+from scanpy._compat import SpBase
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+
+def _as_numpy(x):
+    if hasattr(x, "get"):
+        x = x.get()
+    if isinstance(x, SpBase):
+        return x.toarray()
+    return np.asarray(x)
+
+
+def _conformance_adata() -> AnnData:
+    x = np.array(
+        [
+            [10, 0, 0, 8],
+            [9, 1, 0, 7],
+            [0, 8, 9, 0],
+            [1, 7, 8, 0],
+            [5, 5, 1, 1],
+            [4, 4, 2, 2],
+        ],
+        dtype=np.float32,
+    )
+    return AnnData(x)
+
+
+def _test_normalize_total(backend_name: str) -> None:
+    expected = _conformance_adata()
+    actual = expected.copy()
+
+    sc.pp.normalize_total(expected, target_sum=10, backend="cpu")
+    sc.pp.normalize_total(actual, target_sum=10, backend=backend_name)
+
+    np.testing.assert_allclose(_as_numpy(actual.X), _as_numpy(expected.X))
+
+
+def _test_neighbors(backend_name: str) -> None:
+    expected = _conformance_adata()
+    actual = expected.copy()
+
+    sc.pp.neighbors(expected, n_neighbors=3, use_rep="X", backend="cpu")
+    sc.pp.neighbors(actual, n_neighbors=3, use_rep="X", backend=backend_name)
+
+    for key in ["distances", "connectivities"]:
+        np.testing.assert_allclose(
+            _as_numpy(actual.obsp[key]),
+            _as_numpy(expected.obsp[key]),
+            rtol=1e-5,
+            atol=1e-6,
+        )
+
+
+def _test_umap(backend_name: str) -> None:
+    adata = _conformance_adata()
+    sc.pp.neighbors(adata, n_neighbors=3, use_rep="X", backend="cpu")
+
+    sc.tl.umap(adata, rng=0, backend=backend_name)
+
+    assert "X_umap" in adata.obsm
+    assert adata.obsm["X_umap"].shape == (adata.n_obs, 2)
+    assert np.isfinite(_as_numpy(adata.obsm["X_umap"])).all()
+
+
+_TESTS = {
+    "normalize_total": _test_normalize_total,
+    "neighbors": _test_neighbors,
+    "umap": _test_umap,
+}
+
+
+def validate_backend(
+    backend_name: str,
+    *,
+    functions: Sequence[str] | None = None,
+    raise_on_failure: bool = True,
+) -> dict[str, str]:
+    """Run Scanpy's backend conformance checks against an installed backend."""
+    return run_conformance(
+        backend_name=backend_name,
+        tests=_TESTS,
+        get_backend=get_backend,
+        functions=functions,
+        raise_on_failure=raise_on_failure,
+    )
+
+
+__all__ = ["validate_backend"]
diff --git a/tests/test_backends.py b/tests/test_backends.py
new file mode 100644
index 0000000000..5f24c0edcd
--- /dev/null
+++ b/tests/test_backends.py
@@ -0,0 +1,180 @@
+from __future__ import annotations
+
+from copy import copy
+from inspect import Parameter, signature
+
+import numpy as np
+import pytest
+from anndata import AnnData
+
+import scanpy as sc
+from scanpy._backends import dispatcher
+from scanpy._backends import settings as backend_settings
+from scanpy.testing import validate_backend
+
+
+class FakeRapidsBackend:
+    name = "rapids_singlecell"
+    aliases = ("cuda", "rapids", "rapids-singlecell")
+
+    def normalize_total(
+        self,
+        adata: AnnData,
+        *,
+        target_sum: float | None = None,
+        fake_param: str | None = None,
+    ) -> None:
+        """Fake normalize_total backend implementation.
+
+        Parameters
+        ----------
+        adata
+            Annotated data matrix.
+        target_sum
+            Target total counts.
+        fake_param
+            Backend-only parameter used by tests.
+        """
+        sc.pp.normalize_total(adata, target_sum=target_sum, backend="cpu")
+        adata.uns["fake_backend_called"] = {
+            "function": "normalize_total",
+            "fake_param": fake_param,
+            "target_sum": target_sum,
+        }
+
+
+DISPATCHED_FUNCTIONS = [
+    sc.pp.calculate_qc_metrics,
+    sc.pp.filter_cells,
+    sc.pp.filter_genes,
+    sc.pp.log1p,
+    sc.pp.highly_variable_genes,
+    sc.pp.normalize_total,
+    sc.pp.regress_out,
+    sc.pp.scale,
+    sc.pp.pca,
+    sc.pp.harmony_integrate,
+    sc.pp.scrublet,
+    sc.pp.scrublet_simulate_doublets,
+    sc.pp.neighbors,
+    sc.tl.umap,
+    sc.tl.tsne,
+    sc.tl.diffmap,
+    sc.tl.draw_graph,
+    sc.tl.embedding_density,
+    sc.tl.louvain,
+    sc.tl.leiden,
+    sc.tl.score_genes,
+    sc.tl.score_genes_cell_cycle,
+    sc.tl.rank_genes_groups,
+    sc.get.aggregate,
+]
+
+
+@pytest.fixture
+def fake_rapids_backend():
+    registry = dispatcher._registry
+    dispatch_impl = dispatcher._dispatch_impl
+    old_backend = backend_settings.backend
+    old_state = {
+        "_backends": copy(registry._backends),
+        "_alias_map": copy(registry._alias_map),
+        "_load_errors": copy(registry._load_errors),
+        "_registration_errors": copy(registry._registration_errors),
+        "_warned_untrusted": copy(registry._warned_untrusted),
+        "_discovered": registry._discovered,
+        "_sig_cache": copy(dispatch_impl._sig_cache),
+    }
+
+    backend_settings._backend_var.set("cpu")
+    registry._backends.clear()
+    registry._alias_map.clear()
+    registry._load_errors.clear()
+    registry._registration_errors.clear()
+    registry._warned_untrusted.clear()
+    registry._discovered = True
+    registry._register_backend(
+        FakeRapidsBackend(),
+        entrypoint_name="rapids_singlecell",
+        distribution_name="rapids-singlecell",
+        object_ref="rapids_singlecell.backends.scanpy:ScanpyBackend",
+    )
+    dispatch_impl._sig_cache.clear()
+    dispatch_impl._update_signatures()
+
+    yield
+
+    backend_settings._backend_var.set(old_backend)
+    registry._backends.clear()
+    registry._backends.update(old_state["_backends"])
+    registry._alias_map.clear()
+    registry._alias_map.update(old_state["_alias_map"])
+    registry._load_errors.clear()
+    registry._load_errors.update(old_state["_load_errors"])
+    registry._registration_errors.clear()
+    registry._registration_errors.update(old_state["_registration_errors"])
+    registry._warned_untrusted.clear()
+    registry._warned_untrusted.update(old_state["_warned_untrusted"])
+    registry._discovered = old_state["_discovered"]
+    dispatch_impl._sig_cache.clear()
+    dispatch_impl._sig_cache.update(old_state["_sig_cache"])
+    dispatch_impl._update_signatures()
+
+
+@pytest.mark.parametrize("func", DISPATCHED_FUNCTIONS)
+def test_dispatched_functions_have_backend_keyword(func):
+    backend = signature(func).parameters["backend"]
+
+    assert backend.kind is Parameter.KEYWORD_ONLY
+    assert backend.default is None
+
+
+def test_settings_resolve_rapids_alias(fake_rapids_backend):
+    sc.settings.backend = "cuda"
+
+    assert sc.settings.backend == "rapids_singlecell"
+    assert sc.settings.get_backend("rapids") is not None
+    assert sc.settings.available_backends() == ["rapids_singlecell"]
+
+
+def test_use_backend_dispatches_and_restores(fake_rapids_backend):
+    adata = AnnData(np.ones((2, 2), dtype=np.float32))
+
+    with sc.settings.use_backend("rapids"):
+        assert sc.settings.backend == "rapids_singlecell"
+        sc.pp.normalize_total(adata, target_sum=4, fake_param="from-backend")
+
+    assert sc.settings.backend == "cpu"
+    assert adata.uns["fake_backend_called"] == {
+        "function": "normalize_total",
+        "fake_param": "from-backend",
+        "target_sum": 4,
+    }
+
+
+def test_call_backend_overrides_settings(fake_rapids_backend):
+    adata = AnnData(np.array([[1, 1], [2, 2]], dtype=np.float32))
+
+    with sc.settings.use_backend("cuda"):
+        sc.pp.normalize_total(adata, target_sum=1, backend="cpu")
+
+    assert "fake_backend_called" not in adata.uns
+    np.testing.assert_allclose(adata.X.sum(axis=1), [1, 1])
+
+
+def test_backend_only_parameters_are_injected(fake_rapids_backend):
+    fake_param = signature(sc.pp.normalize_total).parameters["fake_param"]
+
+    assert fake_param.kind is Parameter.KEYWORD_ONLY
+    assert fake_param.default is None
+
+
+def test_backend_conformance_harness(fake_rapids_backend):
+    assert validate_backend("cuda", functions=["normalize_total"]) == {
+        "normalize_total": "PASSED",
+    }
+
+
+def test_reserved_gpu_backend_name():
+    with pytest.raises(ValueError, match="reserved by scanpy"):
+        sc.settings.backend = "gpu"