perf: short-circuit duck-array dispatch helpers for numpy

doc/whats-new.rst

@@ -42,6 +42,12 @@ Documentation
 Internal Changes
 ~~~~~~~~~~~~~~~~
 
+- Speed up :py:meth:`Dataset.load` and other ``.load()`` paths on datasets
+  with many variables by short-circuiting ``is_chunked_array`` for
+  ``numpy.ndarray`` inputs and removing a duplicate ``is_duck_array``
+  call inside the function. Roughly 1.5× faster ``isel().load()`` on a
+  400-scalar-var dataset (0.52 ms → 0.34 ms); no effect on chunked paths.
+
 
 .. _whats-new.2026.04.0:
 

xarray/namedarray/pycompat.py

@@ -8,7 +8,11 @@
 from packaging.version import Version
 
 from xarray.core.utils import is_scalar
-from xarray.namedarray.utils import is_duck_array, is_duck_dask_array
+from xarray.namedarray.utils import (
+    is_dask_collection,
+    is_duck_array,
+    is_duck_dask_array,
+)
 
 integer_types = (int, np.integer)
 
@@ -89,7 +93,16 @@ def mod_version(mod: ModType) -> Version:
 
 
 def is_chunked_array(x: duckarray[Any, Any]) -> bool:
-    return is_duck_dask_array(x) or (is_duck_array(x) and hasattr(x, "chunks"))
+    # Fast path: ndarrays (and ndarray subclasses without a `chunks` attribute,
+    # e.g. MaskedArray) are the dominant non-dask case and skip the duck-array
+    # protocol work entirely. Subclasses that *do* expose `chunks` — test fakes
+    # like DummyChunkedArray, or third-party chunked-array impls — fall through
+    # to the full check below.
+    if isinstance(x, np.ndarray) and not hasattr(x, "chunks"):
+        return False
+    if not is_duck_array(x):
+        return False
+    return is_dask_collection(x) or hasattr(x, "chunks")
 
 
 def is_0d_dask_array(x: duckarray[Any, Any]) -> bool: