TST: run tests on CPU+GPU (data-apis#221)

* TST: run tests on CPU+GPU

* lock

* Simplify conftest

* capitalisation

Author: crusaderky

pixi.lock

@@ -238,7 +238,7 @@ disable_error_code = ["redundant-expr", "unreachable", "no-any-return"]
 
 [[tool.mypy.overrides]]
 # slow/unavailable on Windows; do not add to the lint env
-module = ["dask.*", "jax.*"]
+module = ["dask.*", "jax.*", "torch.*"]
 ignore_missing_imports = true
 
 # pyright

pyproject.toml

@@ -8,10 +8,12 @@
 from types import ModuleType
 from typing import TYPE_CHECKING, ClassVar, cast
 
+from ._utils import _compat
 from ._utils._compat import (
     array_namespace,
     is_dask_array,
     is_jax_array,
+    is_torch_array,
     is_writeable_array,
 )
 from ._utils._helpers import meta_namespace
@@ -298,7 +300,7 @@ def _op(
             and idx.dtype == xp.bool
             and idx.shape == x.shape
         ):
-            y_xp = xp.asarray(y, dtype=x.dtype)
+            y_xp = xp.asarray(y, dtype=x.dtype, device=_compat.device(x))
             if y_xp.ndim == 0:
                 if out_of_place_op:  # add(), subtract(), ...
                     # suppress inf warnings on Dask
@@ -344,6 +346,12 @@ def _op(
             msg = f"Can't update read-only array {x}"
             raise ValueError(msg)
 
+        # Work around bug in PyTorch where __setitem__ doesn't
+        # always support mismatched dtypes
+        # https://github.com/pytorch/pytorch/issues/150017
+        if is_torch_array(y):
+            y = xp.astype(y, x.dtype, copy=False)
+
         # Backends without boolean indexing (other than JAX) crash here
         if in_place_op:  # add(), subtract(), ...
             x[idx] = in_place_op(x[idx], y)

src/array_api_extra/_lib/_at.py

@@ -34,9 +34,11 @@ class Backend(Enum):  # numpydoc ignore=PR01,PR02  # type: ignore[no-subclass-an
     NUMPY_READONLY = "numpy:readonly", _compat.is_numpy_namespace
     CUPY = "cupy", _compat.is_cupy_namespace
     TORCH = "torch", _compat.is_torch_namespace
+    TORCH_GPU = "torch:gpu", _compat.is_torch_namespace
     DASK = "dask.array", _compat.is_dask_namespace
     SPARSE = "sparse", _compat.is_pydata_sparse_namespace
     JAX = "jax.numpy", _compat.is_jax_namespace
+    JAX_GPU = "jax.numpy:gpu", _compat.is_jax_namespace
 
     def __new__(
         cls, value: str, _is_namespace: Callable[[ModuleType], bool]
@@ -54,7 +56,9 @@ def __init__(
 
     def __str__(self) -> str:  # type: ignore[explicit-override]  # pyright: ignore[reportImplicitOverride]  # numpydoc ignore=RT01
         """Pretty-print parameterized test names."""
-        return self.name.lower()
+        return (
+            self.name.lower().replace("_gpu", ":gpu").replace("_readonly", ":readonly")
+        )
 
     @property
     def modname(self) -> str:  # numpydoc ignore=RT01

src/array_api_extra/_lib/_backends.py

@@ -65,6 +65,24 @@ def _check_ns_shape_dtype(
     return desired_xp
 
 
+def _prepare_for_test(array: Array, xp: ModuleType) -> Array:
+    """
+    Ensure that the array can be compared with xp.testing or np.testing.
+
+    This involves transferring it from GPU to CPU memory, densifying it, etc.
+    """
+    if is_torch_namespace(xp):
+        return array.cpu()  # type: ignore[attr-defined]  # pyright: ignore[reportAttributeAccessIssue]
+    if is_pydata_sparse_namespace(xp):
+        return array.todense()  # type: ignore[attr-defined]  # pyright: ignore[reportAttributeAccessIssue]
+    if is_array_api_strict_namespace(xp):
+        # Note: we deliberately did not add a `.to_device` method in _typing.pyi
+        # even if it is required by the standard as many backends don't support it
+        return array.to_device(xp.Device("CPU_DEVICE"))  # type: ignore[attr-defined]  # pyright: ignore[reportAttributeAccessIssue]
+    # Note: nothing to do for CuPy, because it uses a bespoke test function
+    return array
+
+
 def xp_assert_equal(actual: Array, desired: Array, err_msg: str = "") -> None:
     """
     Array-API compatible version of `np.testing.assert_array_equal`.
@@ -84,6 +102,8 @@ def xp_assert_equal(actual: Array, desired: Array, err_msg: str = "") -> None:
     numpy.testing.assert_array_equal : Similar function for NumPy arrays.
     """
     xp = _check_ns_shape_dtype(actual, desired)
+    actual = _prepare_for_test(actual, xp)
+    desired = _prepare_for_test(desired, xp)
 
     if is_cupy_namespace(xp):
         xp.testing.assert_array_equal(actual, desired, err_msg=err_msg)
@@ -102,22 +122,7 @@ def xp_assert_equal(actual: Array, desired: Array, err_msg: str = "") -> None:
     else:
         import numpy as np  # pylint: disable=import-outside-toplevel
 
-        if is_pydata_sparse_namespace(xp):
-            actual = actual.todense()  # type: ignore[attr-defined]  # pyright: ignore[reportAttributeAccessIssue]
-            desired = desired.todense()  # type: ignore[attr-defined]  # pyright: ignore[reportAttributeAccessIssue]
-
-        actual_np = None
-        desired_np = None
-        if is_array_api_strict_namespace(xp):
-            # __array__ doesn't work on array-api-strict device arrays
-            # We need to convert to the CPU device first
-            actual_np = np.asarray(xp.asarray(actual, device=xp.Device("CPU_DEVICE")))
-            desired_np = np.asarray(xp.asarray(desired, device=xp.Device("CPU_DEVICE")))
-
-        # JAX/Dask arrays work with `np.testing`
-        actual_np = actual if actual_np is None else actual_np
-        desired_np = desired if desired_np is None else desired_np
-        np.testing.assert_array_equal(actual_np, desired_np, err_msg=err_msg)  # pyright: ignore[reportUnknownArgumentType]
+        np.testing.assert_array_equal(actual, desired, err_msg=err_msg)
 
 
 def xp_assert_close(
@@ -165,6 +170,9 @@ def xp_assert_close(
     elif rtol is None:
         rtol = 1e-7
 
+    actual = _prepare_for_test(actual, xp)
+    desired = _prepare_for_test(desired, xp)
+
     if is_cupy_namespace(xp):
         xp.testing.assert_allclose(
             actual, desired, rtol=rtol, atol=atol, err_msg=err_msg
@@ -176,26 +184,11 @@ def xp_assert_close(
     else:
         import numpy as np  # pylint: disable=import-outside-toplevel
 
-        if is_pydata_sparse_namespace(xp):
-            actual = actual.todense()  # type: ignore[attr-defined]  # pyright: ignore[reportAttributeAccessIssue]
-            desired = desired.todense()  # type: ignore[attr-defined]  # pyright: ignore[reportAttributeAccessIssue]
-
-        actual_np = None
-        desired_np = None
-        if is_array_api_strict_namespace(xp):
-            # __array__ doesn't work on array-api-strict device arrays
-            # We need to convert to the CPU device first
-            actual_np = np.asarray(xp.asarray(actual, device=xp.Device("CPU_DEVICE")))
-            desired_np = np.asarray(xp.asarray(desired, device=xp.Device("CPU_DEVICE")))
-
-        # JAX/Dask arrays work with `np.testing`
-        actual_np = actual if actual_np is None else actual_np
-        desired_np = desired if desired_np is None else desired_np
-
+        # JAX/Dask arrays work directly with `np.testing`
         assert isinstance(rtol, float)
-        np.testing.assert_allclose(  # pyright: ignore[reportCallIssue]
-            actual_np,  # type: ignore[arg-type]  # pyright: ignore[reportArgumentType]
-            desired_np,  # type: ignore[arg-type]  # pyright: ignore[reportArgumentType]
+        np.testing.assert_allclose(  # type: ignore[call-overload]  # pyright: ignore[reportCallIssue]
+            actual,  # pyright: ignore[reportArgumentType]
+            desired,  # pyright: ignore[reportArgumentType]
             rtol=rtol,
             atol=atol,
             err_msg=err_msg,

src/array_api_extra/_lib/_testing.py

@@ -139,12 +139,31 @@ def xp(
     # in the global scope of the module containing the test function.
     patch_lazy_xp_functions(request, monkeypatch, xp=xp)
 
-    if library == Backend.JAX:
+    if library.like(Backend.JAX):
         import jax
 
         # suppress unused-ignore to run mypy in -e lint as well as -e dev
         jax.config.update("jax_enable_x64", True)  # type: ignore[no-untyped-call,unused-ignore]
 
+        if library == Backend.JAX_GPU:
+            try:
+                device = jax.devices("cuda")[0]
+            except RuntimeError:
+                pytest.skip("no CUDA device available")
+        else:
+            device = jax.devices("cpu")[0]
+        jax.config.update("jax_default_device", device)
+
+    elif library == Backend.TORCH_GPU:
+        import torch.cuda
+
+        if not torch.cuda.is_available():
+            pytest.skip("no CUDA device available")
+        xp.set_default_device("cuda")
+
+    elif library == Backend.TORCH:  # CPU
+        xp.set_default_device("cpu")
+
     yield xp
 
 

tests/conftest.py

@@ -118,6 +118,9 @@ def assert_copy(
                 pytest.mark.skip_xp_backend(  # test passes when copy=False
                     Backend.JAX, reason="bool mask update with shaped rhs"
                 ),
+                pytest.mark.skip_xp_backend(  # test passes when copy=False
+                    Backend.JAX_GPU, reason="bool mask update with shaped rhs"
+                ),
                 pytest.mark.xfail_xp_backend(
                     Backend.DASK, reason="bool mask update with shaped rhs"
                 ),
@@ -247,14 +250,14 @@ def test_incompatible_dtype(
     idx = xp.asarray([True, False]) if bool_mask else slice(None)
     z = None
 
-    if library is Backend.JAX:
+    if library.like(Backend.JAX):
         if bool_mask:
             z = at_op(x, idx, op, 1.1, copy=copy)
         else:
             with pytest.warns(FutureWarning, match="cannot safely cast"):
                 z = at_op(x, idx, op, 1.1, copy=copy)
 
-    elif library is Backend.DASK:
+    elif library.like(Backend.DASK):
         z = at_op(x, idx, op, 1.1, copy=copy)
 
     elif library.like(Backend.ARRAY_API_STRICT) and op is not _AtOp.SET:
@@ -302,6 +305,9 @@ def test_no_inf_warnings(xp: ModuleType, bool_mask: bool):
                     Backend.NUMPY_READONLY, reason="read-only backend"
                 ),
                 pytest.mark.skip_xp_backend(Backend.JAX, reason="read-only backend"),
+                pytest.mark.skip_xp_backend(
+                    Backend.JAX_GPU, reason="read-only backend"
+                ),
                 pytest.mark.skip_xp_backend(Backend.SPARSE, reason="read-only backend"),
             ],
         ),

tests/test_at.py

@@ -234,7 +234,7 @@ def test_hypothesis(  # type: ignore[explicit-any,decorated-any]
 
         # cupy/cupy#8382
         # https://github.com/jax-ml/jax/issues/26658
-        elements = {"allow_subnormal": library not in (Backend.CUPY, Backend.JAX)}
+        elements = {"allow_subnormal": not library.like(Backend.CUPY, Backend.JAX)}
 
         fill_value = xp.asarray(
             data.draw(npst.arrays(dtype=dtype, shape=(), elements=elements))
@@ -930,6 +930,9 @@ class TestSetDiff1D:
     @pytest.mark.xfail_xp_backend(
         Backend.TORCH, reason="index_select not implemented for uint32"
     )
+    @pytest.mark.xfail_xp_backend(
+        Backend.TORCH_GPU, reason="index_select not implemented for uint32"
+    )
     def test_setdiff1d(self, xp: ModuleType):
         x1 = xp.asarray([6, 5, 4, 7, 1, 2, 7, 4])
         x2 = xp.asarray([2, 4, 3, 3, 2, 1, 5])

tests/test_funcs.py

@@ -27,6 +27,9 @@
             True,
             marks=[
                 pytest.mark.skip_xp_backend(Backend.CUPY, reason="device->host copy"),
+                pytest.mark.skip_xp_backend(
+                    Backend.TORCH_GPU, reason="device->host copy"
+                ),
                 pytest.mark.skip_xp_backend(Backend.SPARSE, reason="densification"),
             ],
         ),
@@ -100,6 +103,9 @@ def f(x: Array) -> tuple[Array, Array]:
             True,
             marks=[
                 pytest.mark.skip_xp_backend(Backend.CUPY, reason="device->host copy"),
+                pytest.mark.skip_xp_backend(
+                    Backend.TORCH_GPU, reason="device->host copy"
+                ),
                 pytest.mark.skip_xp_backend(Backend.SPARSE, reason="densification"),
             ],
         ),
@@ -216,7 +222,7 @@ def test_lazy_apply_none_shape_in_args(xp: ModuleType, library: Backend):
     int_type = xp.asarray(0).dtype
 
     ctx: contextlib.AbstractContextManager[object]
-    if library is Backend.JAX:
+    if library.like(Backend.JAX):
         ctx = pytest.raises(ValueError, match="Output shape must be fully known")
     elif library is Backend.ARRAY_API_STRICTEST:
         ctx = pytest.raises(RuntimeError, match="data-dependent shapes")
@@ -254,6 +260,7 @@ def f(x: Array) -> Array:
 
 @pytest.mark.skip_xp_backend(Backend.SPARSE, reason="index by sparse array")
 @pytest.mark.skip_xp_backend(Backend.JAX, reason="boolean indexing")
+@pytest.mark.skip_xp_backend(Backend.JAX_GPU, reason="boolean indexing")
 @pytest.mark.skip_xp_backend(Backend.ARRAY_API_STRICTEST, reason="boolean indexing")
 def test_lazy_apply_none_shape_broadcast(xp: ModuleType):
     """Broadcast from input array with unknown shape"""
@@ -273,6 +280,9 @@ def test_lazy_apply_none_shape_broadcast(xp: ModuleType):
                     Backend.ARRAY_API_STRICT, reason="device->host copy"
                 ),
                 pytest.mark.skip_xp_backend(Backend.CUPY, reason="device->host copy"),
+                pytest.mark.skip_xp_backend(
+                    Backend.TORCH_GPU, reason="device->host copy"
+                ),
                 pytest.mark.skip_xp_backend(Backend.SPARSE, reason="densification"),
             ],
         ),
@@ -321,7 +331,7 @@ def f(x: Array, y: None, z: int | Array) -> Array:
         assert isinstance(x, mtyp)
         assert y is None
         # jax.pure_callback wraps scalar args
-        assert isinstance(z, mtyp if library is Backend.JAX else int)
+        assert isinstance(z, mtyp if library.like(Backend.JAX) else int)
         return x + z
 
     x = xp.asarray([1, 2])

tests/test_lazy.py

@@ -218,6 +218,7 @@ def test_lazy_xp_function_static_params(xp: ModuleType, func: Callable[..., Arra
     erf = None
 
 
+@pytest.mark.skip_xp_backend(Backend.TORCH_GPU, reason="device->host copy")
 @pytest.mark.filterwarnings("ignore:__array_wrap__:DeprecationWarning")  # PyTorch
 def test_lazy_xp_function_cython_ufuncs(xp: ModuleType, library: Backend):
     pytest.importorskip("scipy")
@@ -293,12 +294,12 @@ def test_lazy_xp_modules(xp: ModuleType, library: Backend):
     y = naked.f(x)
     xp_assert_equal(y, x)
 
-    if library is Backend.JAX:
+    if library.like(Backend.JAX):
         with pytest.raises(
             TypeError, match="Attempted boolean conversion of traced array"
         ):
             wrapped.f(x)
-    elif library is Backend.DASK:
+    elif library.like(Backend.DASK):
         with pytest.raises(AssertionError, match=r"dask\.compute"):
             wrapped.f(x)
     else: