[array API] add device property & to_device method

CHANGELOG.md

@@ -25,6 +25,8 @@ Remember to align the itemized text with the first line of an item within a list
     will be removed in a future release.
   * Updated the repr of gpu devices to be more consistent
     with TPUs/CPUs. For example, `cuda(id=0)` will now be `CudaDevice(id=0)`.
+  * Added the `device` property and `to_device` method to {class}`jax.Array`, as
+    part of JAX's [Array API](https://data-apis.org/array-api) support.
 * Deprecations
   * Removed a number of previously-deprecated internal APIs related to
     polymorphic shapes. From {mod}`jax.core`: removed `canonicalize_shape`,

jax/_src/array.py

@@ -254,6 +254,13 @@ def size(self):
   def sharding(self):
     return self._sharding
 
+  @property
+  def device(self):
+    self._check_if_deleted()
+    if isinstance(self.sharding, SingleDeviceSharding):
+      return list(self.sharding.device_set)[0]
+    return self.sharding
+
   @property
   def weak_type(self):
     return self.aval.weak_type

jax/_src/basearray.py

@@ -22,6 +22,7 @@
 from collections.abc import Sequence
 
 # TODO(jakevdp): fix import cycles and define these.
+Device = Any
 Shard = Any
 Sharding = Any
 
@@ -112,6 +113,15 @@ def is_fully_replicated(self) -> bool:
   def sharding(self) -> Sharding:
     """The sharding for the array."""
 
+  @property
+  @abc.abstractmethod
+  def device(self) -> Device | Sharding:
+    """Array API-compatible device attribute.
+
+    For single-device arrays, this returns a Device. For sharded arrays, this
+    returns a Sharding.
+    """
+
 
 Array.__module__ = "jax"
 

jax/_src/basearray.pyi

@@ -204,6 +204,8 @@ class Array(abc.ABC):
   @property
   def sharding(self) -> Sharding: ...
   @property
+  def device(self) -> Device | Sharding: ...
+  @property
   def addressable_shards(self) -> Sequence[Shard]: ...
   @property
   def global_shards(self) -> Sequence[Shard]: ...
@@ -216,6 +218,7 @@ class Array(abc.ABC):
   @property
   def traceback(self) -> Traceback: ...
   def unsafe_buffer_pointer(self) -> int: ...
+  def to_device(self, device: Device | Sharding, *, stream: int | Any | None) -> Array: ...
 
 
 StaticScalar = Union[

jax/_src/core.py

@@ -738,6 +738,15 @@ def sharding(self):
       f"The 'sharding' attribute is not available on {self._error_repr()}."
       f"{self._origin_msg()}")
 
+  @property
+  def device(self):
+    # This attribute is part of the jax.Array API, but only defined on concrete arrays.
+    # Raising a ConcretizationTypeError would make sense, but for backward compatibility
+    # we raise an AttributeError so that hasattr() and getattr() work as expected.
+    raise AttributeError(self,
+      f"The 'device' attribute is not available on {self._error_repr()}."
+      f"{self._origin_msg()}")
+
   @property
   def addressable_shards(self):
     raise ConcretizationTypeError(self,

jax/_src/earray.py

@@ -83,6 +83,12 @@ def sharding(self):
     phys_sharding = self._data.sharding
     return sharding_impls.logical_sharding(self.aval, phys_sharding)
 
+  @property
+  def device(self):
+    if isinstance(self._data.sharding, sharding_impls.SingleDeviceSharding):
+      return self._data.device
+    return self.sharding
+
   # TODO(mattjj): not implemented below here, need more methods from ArrayImpl
 
   def addressable_data(self, index: int) -> EArray:

jax/_src/numpy/array_methods.py

@@ -32,6 +32,7 @@
 import jax
 from jax import lax
 from jax.sharding import Sharding
+from jax._src import api
 from jax._src import core
 from jax._src import dtypes
 from jax._src.api_util import _ensure_index_tuple
@@ -67,6 +68,12 @@ def _astype(arr: ArrayLike, dtype: DTypeLike, copy: bool = False, device: xc.Dev
   """
   return lax_numpy.astype(arr, dtype, copy=copy, device=device)
 
+def _to_device(arr: ArrayLike, device: xc.Device | Sharding, *,
+               stream: int | Any | None = None):
+  if stream is not None:
+    raise NotImplementedError("stream argument of array.to_device()")
+  return api.device_put(arr, device)
+
 
 def _nbytes(arr: ArrayLike) -> int:
   """Total bytes consumed by the elements of the array."""
@@ -694,6 +701,7 @@ def max(self, values, *, indices_are_sorted=False, unique_indices=False,
   "sum": reductions.sum,
   "swapaxes": lax_numpy.swapaxes,
   "take": lax_numpy.take,
+  "to_device": _to_device,
   "trace": lax_numpy.trace,
   "transpose": _transpose,
   "var": reductions.var,

jax/experimental/array_api/_array_methods.py

@@ -31,15 +31,6 @@ def _array_namespace(self, /, *, api_version: None | str = None):
   return jax.experimental.array_api
 
 
-def _to_device(self, device: xe.Device | Sharding | None, *,
-               stream: int | Any | None = None):
-  if stream is not None:
-    raise NotImplementedError("stream argument of array.to_device()")
-  return jax.device_put(self, device)
-
-
 def add_array_object_methods():
   # TODO(jakevdp): set on tracers as well?
   setattr(ArrayImpl, "__array_namespace__", _array_namespace)
-  setattr(ArrayImpl, "to_device", _to_device)
-  setattr(ArrayImpl, "device", property(lambda self: self.sharding))

tests/array_test.py

@@ -1276,6 +1276,30 @@ def test_gspmd_sharding_hash_eq(self):
     self.assertEqual(x1, x2)
     self.assertEqual(hash(x1), hash(x2))
 
+  def test_device_attr(self):
+    # For single-device arrays, x.device returns the device
+    x = jnp.ones((2, 10))
+    self.assertEqual(x.device, list(x.devices())[0])
+
+    # For sharded arrays, x.device returns the sharding
+    mesh = jtu.create_global_mesh((2,), ('x',))
+    sharding = jax.sharding.NamedSharding(mesh, P('x'))
+    x = jax.device_put(x, sharding)
+    self.assertEqual(x.device, sharding)
+
+  def test_to_device(self):
+    device = jax.devices()[-1]
+    mesh = jtu.create_global_mesh((2,), ('x',))
+    sharding = jax.sharding.NamedSharding(mesh, P('x'))
+
+    x = jnp.ones((2, 10))
+
+    x_device = x.to_device(device)
+    x_sharding = x.to_device(sharding)
+
+    self.assertEqual(x_device.device, device)
+    self.assertEqual(x_sharding.device, sharding)
+
 
 class RngShardingTest(jtu.JaxTestCase):
   # tests that the PRNGs are automatically sharded as expected