ENH: CuPy creation functions to respect device= parameter

Author: crusaderky

Diff for: array_api_compat/common/_aliases.py

@@ -7,7 +7,7 @@
 import inspect
 from typing import TYPE_CHECKING, Any, NamedTuple, Optional, Sequence, cast
 
-from ._helpers import _check_device, array_namespace
+from ._helpers import _device_ctx, array_namespace
 from ._helpers import device as _get_device
 from ._helpers import is_cupy_namespace as _is_cupy_namespace
 from ._typing import Array, Device, DType, Namespace
@@ -32,8 +32,8 @@ def arange(
     device: Device | None = None,
     **kwargs: object,
 ) -> Array:
-    _check_device(xp, device)
-    return xp.arange(start, stop=stop, step=step, dtype=dtype, **kwargs)
+    with _device_ctx(xp, device):
+        return xp.arange(start, stop=stop, step=step, dtype=dtype, **kwargs)
 
 
 def empty(
@@ -44,8 +44,8 @@ def empty(
     device: Device | None = None,
     **kwargs: object,
 ) -> Array:
-    _check_device(xp, device)
-    return xp.empty(shape, dtype=dtype, **kwargs)
+    with _device_ctx(xp, device):
+        return xp.empty(shape, dtype=dtype, **kwargs)
 
 
 def empty_like(
@@ -57,8 +57,8 @@ def empty_like(
     device: Device | None = None,
     **kwargs: object,
 ) -> Array:
-    _check_device(xp, device)
-    return xp.empty_like(x, dtype=dtype, **kwargs)
+    with _device_ctx(xp, device, like=x):
+        return xp.empty_like(x, dtype=dtype, **kwargs)
 
 
 def eye(
@@ -72,8 +72,8 @@ def eye(
     device: Device | None = None,
     **kwargs: object,
 ) -> Array:
-    _check_device(xp, device)
-    return xp.eye(n_rows, M=n_cols, k=k, dtype=dtype, **kwargs)
+    with _device_ctx(xp, device):
+        return xp.eye(n_rows, M=n_cols, k=k, dtype=dtype, **kwargs)
 
 
 def full(
@@ -85,8 +85,8 @@ def full(
     device: Device | None = None,
     **kwargs: object,
 ) -> Array:
-    _check_device(xp, device)
-    return xp.full(shape, fill_value, dtype=dtype, **kwargs)
+    with _device_ctx(xp, device):
+        return xp.full(shape, fill_value, dtype=dtype, **kwargs)
 
 
 def full_like(
@@ -99,8 +99,8 @@ def full_like(
     device: Device | None = None,
     **kwargs: object,
 ) -> Array:
-    _check_device(xp, device)
-    return xp.full_like(x, fill_value, dtype=dtype, **kwargs)
+    with _device_ctx(xp, device, like=x):
+        return xp.full_like(x, fill_value, dtype=dtype, **kwargs)
 
 
 def linspace(
@@ -115,8 +115,8 @@ def linspace(
     endpoint: bool = True,
     **kwargs: object,
 ) -> Array:
-    _check_device(xp, device)
-    return xp.linspace(start, stop, num, dtype=dtype, endpoint=endpoint, **kwargs)
+    with _device_ctx(xp, device):
+        return xp.linspace(start, stop, num, dtype=dtype, endpoint=endpoint, **kwargs)
 
 
 def ones(
@@ -127,8 +127,8 @@ def ones(
     device: Device | None = None,
     **kwargs: object,
 ) -> Array:
-    _check_device(xp, device)
-    return xp.ones(shape, dtype=dtype, **kwargs)
+    with _device_ctx(xp, device):
+        return xp.ones(shape, dtype=dtype, **kwargs)
 
 
 def ones_like(
@@ -140,8 +140,8 @@ def ones_like(
     device: Device | None = None,
     **kwargs: object,
 ) -> Array:
-    _check_device(xp, device)
-    return xp.ones_like(x, dtype=dtype, **kwargs)
+    with _device_ctx(xp, device, like=x):
+        return xp.ones_like(x, dtype=dtype, **kwargs)
 
 
 def zeros(
@@ -152,8 +152,8 @@ def zeros(
     device: Device | None = None,
     **kwargs: object,
 ) -> Array:
-    _check_device(xp, device)
-    return xp.zeros(shape, dtype=dtype, **kwargs)
+    with _device_ctx(xp, device):
+        return xp.zeros(shape, dtype=dtype, **kwargs)
 
 
 def zeros_like(
@@ -165,8 +165,8 @@ def zeros_like(
     device: Device | None = None,
     **kwargs: object,
 ) -> Array:
-    _check_device(xp, device)
-    return xp.zeros_like(x, dtype=dtype, **kwargs)
+    with _device_ctx(xp, device, like=x):
+        return xp.zeros_like(x, dtype=dtype, **kwargs)
 
 
 # np.unique() is split into four functions in the array API:

Diff for: array_api_compat/common/_helpers.py

@@ -8,11 +8,12 @@
 
 from __future__ import annotations
 
+import contextlib
 import inspect
 import math
 import sys
 import warnings
-from collections.abc import Collection
+from collections.abc import Collection, Generator
 from typing import (
     TYPE_CHECKING,
     Any,
@@ -663,26 +664,42 @@ def your_function(x, y):
 get_namespace = array_namespace
 
 
-def _check_device(bare_xp: Namespace, device: Device) -> None:  # pyright: ignore[reportUnusedFunction]
-    """
-    Validate dummy device on device-less array backends.
-
-    Notes
-    -----
-    This function is also invoked by CuPy, which does have multiple devices
-    if there are multiple GPUs available.
-    However, CuPy multi-device support is currently impossible
-    without using the global device or a context manager:
+def _device_ctx(
+    bare_xp: Namespace, device: Device, like: Array | None = None
+) -> Generator[None]:
+    """Context manager which changes the current device in CuPy.
 
-    https://github.com/data-apis/array-api-compat/pull/293
+    Used internally by array creation functions in common._aliases.
     """
-    if bare_xp is sys.modules.get("numpy"):
-        if device not in ("cpu", None):
+    if device is None:
+        if like is None:
+            return contextlib.nullcontext()    
+        device = _device(like)
+
+    if bare_xp is sys.modules.get('numpy'):
+        if device != "cpu":
             raise ValueError(f"Unsupported device for NumPy: {device!r}")
+        return contextlib.nullcontext()
 
-    elif bare_xp is sys.modules.get("dask.array"):
-        if device not in ("cpu", _DASK_DEVICE, None):
+    if bare_xp is sys.modules.get('dask.array'):
+        if device not in ("cpu", _DASK_DEVICE):
             raise ValueError(f"Unsupported device for Dask: {device!r}")
+        return contextlib.nullcontext()
+
+    if bare_xp is sys.modules.get('cupy'):
+        if not isinstance(device, bare_xp.cuda.Device):
+            raise TypeError(f"device is not a cupy.cuda.Device: {device!r}")
+        return device
+
+    # PyTorch doesn't have a "current device" context manager and you
+    # can't use array creation functions from common._aliases.
+    raise AssertionError("unreachable")  # pragma: nocover
+
+
+def _check_device(bare_xp: Namespace, device: Device) -> None:
+    """Validate dummy device on device-less array backends."""
+    with _device_ctx(bare_xp, device):
+        pass
 
 
 # Placeholder object to represent the dask device
@@ -781,42 +798,26 @@ def _cupy_to_device(
     /,
     stream: int | Any | None = None,
 ) -> _CupyArray:
-    import cupy as cp  # pyright: ignore[reportMissingTypeStubs]
-    from cupy.cuda import Device as _Device  # pyright: ignore
-    from cupy.cuda import stream as stream_module  # pyright: ignore
-    from cupy_backends.cuda.api import runtime  # pyright: ignore
+    import cupy as cp
 
-    if device == x.device:
-        return x
-    elif device == "cpu":
+    if device == "cpu":
         # allowing us to use `to_device(x, "cpu")`
         # is useful for portable test swapping between
         # host and device backends
         return x.get()
-    elif not isinstance(device, _Device):
-        raise ValueError(f"Unsupported device {device!r}")
-    else:
-        # see cupy/cupy#5985 for the reason how we handle device/stream here
-        prev_device: Any = runtime.getDevice()  # pyright: ignore[reportUnknownMemberType]
-        prev_stream = None
-        if stream is not None:
-            prev_stream: Any = stream_module.get_current_stream()  # pyright: ignore
-            # stream can be an int as specified in __dlpack__, or a CuPy stream
-            if isinstance(stream, int):
-                stream = cp.cuda.ExternalStream(stream)  # pyright: ignore
-            elif isinstance(stream, cp.cuda.Stream):  # pyright: ignore[reportUnknownMemberType]
-                pass
-            else:
-                raise ValueError("the input stream is not recognized")
-            stream.use()  # pyright: ignore[reportUnknownMemberType]
-        try:
-            runtime.setDevice(device.id)  # pyright: ignore[reportUnknownMemberType]
-            arr = x.copy()
-        finally:
-            runtime.setDevice(prev_device)  # pyright: ignore[reportUnknownMemberType]
-            if stream is not None:
-                prev_stream.use()
-        return arr
+    if not isinstance(device, cp.cuda.Device):
+        raise TypeError(f"Unsupported device {device!r}")
+
+    # stream can be an int as specified in __dlpack__, or a CuPy stream
+    if isinstance(stream, int):
+        stream = cp.cuda.ExternalStream(stream)
+    elif stream is None:
+        stream = contextlib.nullcontext()
+    elif not isinstance(stream, cp.cuda.Stream):
+        raise TypeError('the input stream is not recognized')
+
+    with device, stream:
+        return cp.asarray(x)
 
 
 def _torch_to_device(

Diff for: array_api_compat/cupy/_aliases.py

@@ -64,8 +64,6 @@
 finfo = get_xp(cp)(_aliases.finfo)
 iinfo = get_xp(cp)(_aliases.iinfo)
 
-_copy_default = object()
-
 
 # asarray also adds the copy keyword, which is not present in numpy 1.0.
 def asarray(
@@ -79,7 +77,7 @@ def asarray(
     *,
     dtype: Optional[DType] = None,
     device: Optional[Device] = None,
-    copy: Optional[bool] = _copy_default,
+    copy: Optional[bool] = None,
     **kwargs,
 ) -> Array:
     """
@@ -88,26 +86,15 @@ def asarray(
     See the corresponding documentation in the array library and/or the array API
     specification for more details.
     """
-    with cp.cuda.Device(device):
-        # cupy is like NumPy 1.26 (except without _CopyMode). See the comments
-        # in asarray in numpy/_aliases.py.
-        if copy is not _copy_default:
-            # A future version of CuPy will change the meaning of copy=False
-            # to mean no-copy. We don't know for certain what version it will
-            # be yet, so to avoid breaking that version, we use a different
-            # default value for copy so asarray(obj) with no copy kwarg will
-            # always do the copy-if-needed behavior.
-
-            # This will still need to be updated to remove the
-            # NotImplementedError for copy=False, but at least this won't
-            # break the default or existing behavior.
-            if copy is None:
-                copy = False
-            elif copy is False:
-                raise NotImplementedError("asarray(copy=False) is not yet supported in cupy")
-            kwargs['copy'] = copy
-
-        return cp.array(obj, dtype=dtype, **kwargs)
+    if copy is False:
+        raise NotImplementedError("asarray(copy=False) is not yet supported in cupy")
+
+    like = obj if isinstance(obj, cp.ndarray) else None
+    with _helpers._device_ctx(cp, device, like=like):
+        if copy is None:
+            return cp.asarray(obj, dtype=dtype, **kwargs)
+        else:
+            return cp.array(obj, dtype=dtype, copy=True, **kwargs)
 
 
 def astype(