BUG: .device attribute inside jax.jit

Author: crusaderky

array_api_compat/common/_helpers.py

@@ -658,14 +658,18 @@ def device(x: Array, /) -> Device:
             pass
         return _DASK_DEVICE
     elif is_jax_array(x):
-        # JAX has .device() as a method, but it is being deprecated so that it
-        # can become a property, in accordance with the standard. In order for
-        # this function to not break when JAX makes the flip, we check for
-        # both here.
-        if inspect.ismethod(x.device):
-            return x.device()
+        # FIXME Jitted JAX arrays do not have a device attribute
+        #       https://github.com/jax-ml/jax/issues/26000
+        #       Return None in this case. Note that this workaround breaks
+        #       the standard and will result in new arrays being created on the
+        #       default device instead of the same device as the input array(s).
+        x_device = getattr(x, 'device', None)
+        # Older JAX releases had .device() as a method, which has been replaced
+        # with a property in accordance with the standard.
+        if inspect.ismethod(x_device):
+            return x_device()
         else:
-            return x.device
+            return x_device
     elif is_pydata_sparse_array(x):
         # `sparse` will gain `.device`, so check for this first.
         x_device = getattr(x, 'device', None)
@@ -796,8 +800,11 @@ def to_device(x: Array, device: Device, /, *, stream: Optional[Union[int, Any]]
         raise ValueError(f"Unsupported device {device!r}")
     elif is_jax_array(x):
         if not hasattr(x, "__array_namespace__"):
-            # In JAX v0.4.31 and older, this import adds to_device method to x.
+            # In JAX v0.4.31 and older, this import adds to_device method to x...
             import jax.experimental.array_api # noqa: F401
+            # ... but only on eager JAX. It won't work inside jax.jit.
+            if not hasattr(x, "to_device"):
+                return x
         return x.to_device(device, stream=stream)
     elif is_pydata_sparse_array(x) and device == _device(x):
         # Perform trivial check to return the same array if

tests/test_array_namespace.py

@@ -22,7 +22,7 @@ def test_array_namespace(library, api_version, use_compat):
     if use_compat is True and library in {'array_api_strict', 'jax.numpy', 'sparse'}:
         pytest.raises(ValueError, lambda: array_namespace(array, use_compat=use_compat))
         return
-    namespace = array_api_compat.array_namespace(array, api_version=api_version, use_compat=use_compat)
+    namespace = array_namespace(array, api_version=api_version, use_compat=use_compat)
 
     if use_compat is False or use_compat is None and library not in wrapped_libraries:
         if library == "jax.numpy" and use_compat is None:
@@ -44,7 +44,7 @@ def test_array_namespace(library, api_version, use_compat):
 
     if library == "numpy":
         # check that the same namespace is returned for NumPy scalars
-        scalar_namespace = array_api_compat.array_namespace(
+        scalar_namespace = array_namespace(
             xp.float64(0.0), api_version=api_version, use_compat=use_compat
         )
         assert scalar_namespace == namespace
@@ -75,8 +75,7 @@ def test_array_namespace(library, api_version, use_compat):
 def test_jax_zero_gradient():
     jx = jax.numpy.arange(4)
     jax_zero = jax.vmap(jax.grad(jax.numpy.float32, allow_int=True))(jx)
-    assert (array_api_compat.get_namespace(jax_zero) is
-            array_api_compat.get_namespace(jx))
+    assert array_namespace(jax_zero) is array_namespace(jx)
 
 def test_array_namespace_errors():
     pytest.raises(TypeError, lambda: array_namespace([1]))
@@ -91,7 +90,7 @@ def test_array_namespace_errors_torch():
     x = np.asarray([1, 2])
     pytest.raises(TypeError, lambda: array_namespace(x, y))
 
-def test_api_version():
+def test_api_version_torch():
     x = torch.asarray([1, 2])
     torch_ = import_("torch", wrapper=True)
     assert array_namespace(x, api_version="2023.12") == torch_
@@ -113,7 +112,7 @@ def test_api_version():
 
 def test_get_namespace():
     # Backwards compatible wrapper
-    assert array_api_compat.get_namespace is array_api_compat.array_namespace
+    assert array_api_compat.get_namespace is array_namespace
 
 def test_python_scalars():
     a = torch.asarray([1, 2])

tests/test_common.py

@@ -3,7 +3,7 @@
 import pytest
 import numpy as np
 import array
-from numpy.testing import assert_allclose
+from numpy.testing import assert_equal
 
 from array_api_compat import (  # noqa: F401
     is_numpy_array, is_cupy_array, is_torch_array,
@@ -188,7 +188,10 @@ def test_device(library):
     dev = device(x)
 
     x2 = to_device(x, dev)
-    assert device(x) == device(x2)
+    assert device(x2) == device(x)
+
+    x3 = xp.asarray(x, device=dev)
+    assert device(x3) == device(x)
 
 
 @pytest.mark.parametrize("library", wrapped_libraries)
@@ -207,7 +210,7 @@ def test_to_device_host(library):
     # a `device(x)` query; however, what's really important
     # here is that we can test portably after calling
     # to_device(x, "cpu") to return to host
-    assert_allclose(x, expected)
+    assert_equal(x, expected)
 
 
 @pytest.mark.parametrize("target_library", is_array_functions.keys())

tests/test_jax.py

@@ -0,0 +1,27 @@
+import jax
+import jax.numpy as jnp
+from numpy.testing import assert_equal
+import pytest
+
+from array_api_compat import device, to_device
+
+
+@pytest.mark.parametrize(
+    "func", 
+    [
+        lambda x: jnp.zeros(1, device=device(x)),
+        lambda x: jnp.zeros_like(jnp.ones(1, device=device(x))),
+        lambda x: jnp.zeros_like(jnp.empty(1, device=device(x))),
+        lambda x: jnp.full(1, fill_value=0, device=device(x)),
+        lambda x: jnp.asarray([0], device=device(x)),
+        lambda x: to_device(jnp.zeros(1), device(x)),
+    ]
+)
+def test_device_jit(func):
+    # Test work around to https://github.com/jax-ml/jax/issues/26000
+    # Also test missing to_device() method in JAX < 0.4.31
+    # when inside jax.jit, even after importing jax.experimental.array_api
+
+    x = jnp.ones(1)
+    assert_equal(func(x), jnp.asarray([0]))
+    assert_equal(jax.jit(func)(x), jnp.asarray([0]))