Address feedback

ndonnx/_array.py

@@ -256,12 +256,12 @@ def shape(self) -> tuple[int | None, ...]:
 
     @property
     def device(self):
-        return ndonnx_device
+        return device
 
     def to_device(
-        self, device: NdonnxDevice, /, *, stream: int | Any | None = None
+        self, device: _Device, /, *, stream: int | Any | None = None
     ) -> Array:
-        if device is not ndonnx_device:
+        if device is not device:
             raise ValueError("Cannot move Array to a different device")
         if stream is not None:
             raise ValueError("The 'stream' parameter is not supported in ndonnx.")
@@ -592,19 +592,23 @@ def any(self, axis: int | None = 0, keepdims: bool | None = False) -> ndx.Array:
         return ndx.any(self, axis=axis, keepdims=False)
 
 
-class NdonnxDevice:
+class _Device:
+    # We would rather not give users the impression that their arrays
+    # are tied to a specific device when serializing an ONNX graph as
+    # such a concept does not exist in the ONNX standard.
+
     def __str__(self):
         return "ndonnx device"
 
     def __eq__(self, other):
-        return type(other) is NdonnxDevice
+        return type(other) is _Device
 
 
-ndonnx_device = NdonnxDevice()
+device = _Device()
 
 
 __all__ = [
     "Array",
     "array",
-    "ndonnx_device",
+    "device",
 ]

ndonnx/_info.py

@@ -4,7 +4,7 @@
 from __future__ import annotations
 
 import ndonnx as ndx
-from ndonnx._array import ndonnx_device
+from ndonnx._array import device
 from ndonnx._data_types import canonical_name
 
 
@@ -32,18 +32,14 @@ def capabilities(self) -> dict:
         }
 
     def default_device(self):
-        return ndonnx_device
+        return device
 
     def devices(self) -> list:
-        return [ndonnx_device]
+        return [device]
 
     def dtypes(
         self, *, device=None, kind: str | tuple[str, ...] | None = None
     ) -> dict[str, ndx.CoreType]:
-        # We don't care for device since we are writing ONNX graphs.
-        # We would rather not give users the impression that their arrays
-        # are tied to a specific device when serializing an ONNX graph as
-        # such a concept does not exist in the ONNX standard.
         out: dict[str, ndx.CoreType] = {}
         for dtype in self._all_array_api_types:
             if kind is None or ndx.isdtype(dtype, kind):
@@ -54,15 +50,11 @@ def default_dtypes(
         self,
         *,
         device=None,
-    ) -> dict[str, ndx.CoreType | None]:
-        # See comment in `dtypes` method regarding device.
+    ) -> dict[str, ndx.CoreType]:
         return {
             "real floating": ndx.float64,
             "integral": ndx.int64,
             "indexing": ndx.int64,
-            # We don't support complex numbers yet due to immaturity in the ONNX ecoystem, so "complex floating" is meaningless.
-            # The Array API standard requires this key to be present so we set it to None.
-            "complex floating": None,
         }