scipy 1.15.3__cp313-cp313t-macosx_14_0_arm64.whl → 1.16.0rc2__cp313-cp313t-macosx_14_0_arm64.whl

scipy/_lib/_array_api_compat_vendor.py

@@ -0,0 +1,9 @@
+# DO NOT RENAME THIS FILE
+# This is a hook for array_api_extra/src/array_api_extra/_lib/_compat.py
+# to override functions of array_api_compat.
+
+from .array_api_compat import *  # noqa: F403
+from ._array_api import array_namespace as scipy_array_namespace
+
+# overrides array_api_compat.array_namespace inside array-api-extra
+array_namespace = scipy_array_namespace  # type: ignore[assignment]

scipy/_lib/_bunch.py

@@ -194,6 +194,10 @@ def __setattr__(self, key, val):
         '_asdict': _asdict,
         '_extra_fields': extra_field_names,
         '__getnewargs_ex__': __getnewargs_ex__,
+        # _field_defaults and _replace are added to get Polars to detect
+        # a bunch object as a namedtuple. See gh-22450
+        '_field_defaults': {},
+        '_replace': None,
     }
     for index, name in enumerate(field_names):
 

scipy/_lib/_docscrape.py

@@ -703,7 +703,7 @@ class ClassDoc(NumpyDocString):
                 and not self._should_skip_member(name, self._cls)
                 and (
                     func is None
-                    or isinstance(func, (property, cached_property))
+                    or isinstance(func, property | cached_property)
                     or inspect.isdatadescriptor(func)
                 )
                 and self._is_show_member(name)

scipy/_lib/_elementwise_iterative_method.py

@@ -15,7 +15,8 @@
 import math
 import numpy as np
 from ._util import _RichResult, _call_callback_maybe_halt
-from ._array_api import array_namespace, xp_size
+from ._array_api import array_namespace, xp_size, xp_result_type
+import scipy._lib.array_api_extra as xpx
 
 _ESIGNERR = -1
 _ECONVERR = -2
@@ -82,9 +83,8 @@ def _initialize(func, xs, args, complex_ok=False, preserve_shape=None, xp=None):
     # and cause failure.
     # There might be benefit to combining the `xs` into a single array and
     # calling `func` once on the combined array. For now, keep them separate.
+    xat = xp_result_type(*xs, force_floating=True, xp=xp)
     xas = xp.broadcast_arrays(*xs, *args)  # broadcast and rename
-    xat = xp.result_type(*[xa.dtype for xa in xas])
-    xat = xp.asarray(1.).dtype if xp.isdtype(xat, "integral") else xat
     xs, args = xas[:nx], xas[nx:]
     xs = [xp.asarray(x, dtype=xat) for x in xs]  # use copy=False when implemented
     fs = [xp.asarray(func(x, *args)) for x in xs]
@@ -264,7 +264,7 @@ def _loop(work, callback, shape, maxiter, func, args, dtype, pre_func_eval,
 
         post_termination_check(work)
 
-    work.status[:] = _ECALLBACK if cb_terminate else _ECONVERR
+    work.status = xpx.at(work.status)[:].set(_ECALLBACK if cb_terminate else _ECONVERR)
     return _prepare_result(work, res, res_work_pairs, active, shape,
                            customize_result, preserve_shape, xp)
 
@@ -290,14 +290,10 @@ def _check_termination(work, res, res_work_pairs, active, check_termination,
         if not preserve_shape:
             # compress the arrays to avoid unnecessary computation
             for key, val in work.items():
-                # Need to find a better way than these try/excepts
-                # Somehow need to keep compressible numerical args separate
-                if key == 'args':
+                # `continued_fraction` hacks `n`; improve if this becomes a problem
+                if key in {'args', 'n'}:
                     continue
-                try:
-                    work[key] = val[proceed]
-                except (IndexError, TypeError, KeyError):  # not a compressible array
-                    work[key] = val
+                work[key] = val[proceed] if getattr(val, 'ndim', 0) > 0 else val
             work.args = [arg[proceed] for arg in work.args]
 
     return active
@@ -314,28 +310,21 @@ def _update_active(work, res, res_work_pairs, active, mask, preserve_shape, xp):
     if mask is not None:
         if preserve_shape:
             active_mask = xp.zeros_like(mask)
-            active_mask[active] = 1
+            active_mask = xpx.at(active_mask)[active].set(True)
             active_mask = active_mask & mask
             for key, val in update_dict.items():
-                try:
-                    res[key][active_mask] = val[active_mask]
-                except (IndexError, TypeError, KeyError):
-                    res[key][active_mask] = val
+                val = val[active_mask] if getattr(val, 'ndim', 0) > 0 else val
+                res[key] = xpx.at(res[key])[active_mask].set(val)
         else:
             active_mask = active[mask]
             for key, val in update_dict.items():
-                try:
-                    res[key][active_mask] = val[mask]
-                except (IndexError, TypeError, KeyError):
-                    res[key][active_mask] = val
+                val = val[mask] if getattr(val, 'ndim', 0) > 0 else val
+                res[key] = xpx.at(res[key])[active_mask].set(val)
     else:
         for key, val in update_dict.items():
-            if preserve_shape:
-                try:
-                    val = val[active]
-                except (IndexError, TypeError, KeyError):
-                    pass
-            res[key][active] = val
+            if preserve_shape and getattr(val, 'ndim', 0) > 0:
+                val = val[active]
+            res[key] = xpx.at(res[key])[active].set(val)
 
 
 def _prepare_result(work, res, res_work_pairs, active, shape, customize_result,

scipy/_lib/_sparse.py

@@ -0,0 +1,41 @@
+from abc import ABC
+
+__all__ = ["SparseABC", "issparse"]
+
+
+class SparseABC(ABC):
+    pass
+
+
+def issparse(x):
+    """Is `x` of a sparse array or sparse matrix type?
+
+    Parameters
+    ----------
+    x
+        object to check for being a sparse array or sparse matrix
+
+    Returns
+    -------
+    bool
+        True if `x` is a sparse array or a sparse matrix, False otherwise
+
+    Notes
+    -----
+    Use `isinstance(x, sp.sparse.sparray)` to check between an array or matrix.
+    Use `a.format` to check the sparse format, e.g. `a.format == 'csr'`.
+
+    Examples
+    --------
+    >>> import numpy as np
+    >>> from scipy.sparse import csr_array, csr_matrix, issparse
+    >>> issparse(csr_matrix([[5]]))
+    True
+    >>> issparse(csr_array([[5]]))
+    True
+    >>> issparse(np.array([[5]]))
+    False
+    >>> issparse(5)
+    False
+    """
+    return isinstance(x, SparseABC)

scipy/_lib/_testutils.py

@@ -95,8 +95,12 @@ class PytestTester:
 
         pytest_args = ['--showlocals', '--tb=short']
 
-        if extra_argv:
-            pytest_args += list(extra_argv)
+        if extra_argv is None:
+            extra_argv = []
+        pytest_args += extra_argv
+        if any(arg == "-m" or arg == "--markers" for arg in extra_argv):
+            # Likely conflict with default --mode=fast
+            raise ValueError("Must specify -m before --")
 
         if verbose and int(verbose) > 1:
             pytest_args += ["-" + "v"*(int(verbose)-1)]

scipy/_lib/_util.py

@@ -7,27 +7,18 @@ import numbers
 from collections import namedtuple
 import inspect
 import math
-from typing import TypeAlias, TypeVar
+from types import ModuleType
+from typing import Literal, TypeAlias, TypeVar
 
 import numpy as np
-from scipy._lib._array_api import array_namespace, is_numpy, xp_size
+from scipy._lib._array_api import (Array, array_namespace, is_lazy_array,
+                                   is_numpy, is_marray, xp_result_device,
+                                   xp_size, xp_result_type)
 from scipy._lib._docscrape import FunctionDoc, Parameter
+from scipy._lib._sparse import issparse
 
+from numpy.exceptions import AxisError
 
-AxisError: type[Exception]
-ComplexWarning: type[Warning]
-VisibleDeprecationWarning: type[Warning]
-
-if np.lib.NumpyVersion(np.__version__) >= '1.25.0':
-    from numpy.exceptions import (
-        AxisError, ComplexWarning, VisibleDeprecationWarning,
-        DTypePromotionError
-    )
-else:
-    from numpy import (  # type: ignore[attr-defined, no-redef]
-        AxisError, ComplexWarning, VisibleDeprecationWarning  # noqa: F401
-    )
-    DTypePromotionError = TypeError  # type: ignore
 
 np_long: type
 np_ulong: type
@@ -81,85 +72,6 @@ except ImportError:
         pass
 
 
-def _lazywhere(cond, arrays, f, fillvalue=None, f2=None):
-    """Return elements chosen from two possibilities depending on a condition
-
-    Equivalent to ``f(*arrays) if cond else fillvalue`` performed elementwise.
-
-    Parameters
-    ----------
-    cond : array
-        The condition (expressed as a boolean array).
-    arrays : tuple of array
-        Arguments to `f` (and `f2`). Must be broadcastable with `cond`.
-    f : callable
-        Where `cond` is True, output will be ``f(arr1[cond], arr2[cond], ...)``
-    fillvalue : object
-        If provided, value with which to fill output array where `cond` is
-        not True.
-    f2 : callable
-        If provided, output will be ``f2(arr1[cond], arr2[cond], ...)`` where
-        `cond` is not True.
-
-    Returns
-    -------
-    out : array
-        An array with elements from the output of `f` where `cond` is True
-        and `fillvalue` (or elements from the output of `f2`) elsewhere. The
-        returned array has data type determined by Type Promotion Rules
-        with the output of `f` and `fillvalue` (or the output of `f2`).
-
-    Notes
-    -----
-    ``xp.where(cond, x, fillvalue)`` requires explicitly forming `x` even where
-    `cond` is False. This function evaluates ``f(arr1[cond], arr2[cond], ...)``
-    onle where `cond` ``is True.
-
-    Examples
-    --------
-    >>> import numpy as np
-    >>> a, b = np.array([1, 2, 3, 4]), np.array([5, 6, 7, 8])
-    >>> def f(a, b):
-    ...     return a*b
-    >>> _lazywhere(a > 2, (a, b), f, np.nan)
-    array([ nan,  nan,  21.,  32.])
-
-    """
-    xp = array_namespace(cond, *arrays)
-
-    if (f2 is fillvalue is None) or (f2 is not None and fillvalue is not None):
-        raise ValueError("Exactly one of `fillvalue` or `f2` must be given.")
-
-    args = xp.broadcast_arrays(cond, *arrays)
-    bool_dtype = xp.asarray([True]).dtype  # numpy 1.xx doesn't have `bool`
-    cond, arrays = xp.astype(args[0], bool_dtype, copy=False), args[1:]
-
-    temp1 = xp.asarray(f(*(arr[cond] for arr in arrays)))
-
-    if f2 is None:
-        # If `fillvalue` is a Python scalar and we convert to `xp.asarray`, it gets the
-        # default `int` or `float` type of `xp`, so `result_type` could be wrong.
-        # `result_type` should/will handle mixed array/Python scalars;
-        # remove this special logic when it does.
-        if type(fillvalue) in {bool, int, float, complex}:
-            with np.errstate(invalid='ignore'):
-                dtype = (temp1 * fillvalue).dtype
-        else:
-           dtype = xp.result_type(temp1.dtype, fillvalue)
-        out = xp.full(cond.shape, dtype=dtype,
-                      fill_value=xp.asarray(fillvalue, dtype=dtype))
-    else:
-        ncond = ~cond
-        temp2 = xp.asarray(f2(*(arr[ncond] for arr in arrays)))
-        dtype = xp.result_type(temp1, temp2)
-        out = xp.empty(cond.shape, dtype=dtype)
-        out[ncond] = temp2
-
-    out[cond] = temp1
-
-    return out
-
-
 def _lazyselect(condlist, choicelist, arrays, default=0):
     """
     Mimic `np.select(condlist, choicelist)`.
@@ -524,8 +436,7 @@ def _asarray_validated(a, check_finite=True,
 
     """
     if not sparse_ok:
-        import scipy.sparse
-        if scipy.sparse.issparse(a):
+        if issparse(a):
             msg = ('Sparse arrays/matrices are not supported by this function. '
                    'Perhaps one of the `scipy.sparse.linalg` functions '
                    'would work instead.')
@@ -657,6 +568,32 @@ class _FunctionWrapper:
         return self.f(x, *self.args)
 
 
+class _ScalarFunctionWrapper:
+    """
+    Object to wrap scalar user function, allowing picklability
+    """
+    def __init__(self, f, args=None):
+        self.f = f
+        self.args = [] if args is None else args
+        self.nfev = 0
+
+    def __call__(self, x):
+        # Send a copy because the user may overwrite it.
+        # The user of this class might want `x` to remain unchanged.
+        fx = self.f(np.copy(x), *self.args)
+        self.nfev += 1
+
+        # Make sure the function returns a true scalar
+        if not np.isscalar(fx):
+            try:
+                fx = np.asarray(fx).item()
+            except (TypeError, ValueError) as e:
+                raise ValueError(
+                    "The user-provided objective function "
+                    "must return a scalar value."
+                ) from e
+        return fx
+
 class MapWrapper:
     """
     Parallelisation wrapper for working with map-like callables, such as
@@ -731,6 +668,29 @@ class MapWrapper:
                             " form f(func, iterable)") from e
 
 
+def _workers_wrapper(func):
+    """
+    Wrapper to deal with setup-cleanup of workers outside a user function via a
+    ContextManager. It saves having to do the setup/tear down with within that
+    function, which can be messy.
+    """
+    @functools.wraps(func)
+    def inner(*args, **kwds):
+        kwargs = kwds.copy()
+        if 'workers' not in kwargs:
+            _workers = map
+        elif 'workers' in kwargs and kwargs['workers'] is None:
+            _workers = map
+        else:
+            _workers = kwargs['workers']
+
+        with MapWrapper(_workers) as mf:
+            kwargs['workers'] = mf
+            return func(*args, **kwargs)
+
+    return inner
+
+
 def rng_integers(gen, low, high=None, size=None, dtype='int64',
                  endpoint=False):
     """
@@ -923,25 +883,30 @@ def _nan_allsame(a, axis, keepdims=False):
     return ((a0 == a) | np.isnan(a)).all(axis=axis, keepdims=keepdims)
 
 
-def _contains_nan(a, nan_policy='propagate', policies=None, *,
-                  xp_omit_okay=False, xp=None):
+def _contains_nan(
+    a: Array,
+    nan_policy: Literal["propagate", "raise", "omit"] = "propagate",
+    *,
+    xp_omit_okay: bool = False, 
+    xp: ModuleType | None = None,
+) -> Array | bool:
     # Regarding `xp_omit_okay`: Temporarily, while `_axis_nan_policy` does not
     # handle non-NumPy arrays, most functions that call `_contains_nan` want
     # it to raise an error if `nan_policy='omit'` and `xp` is not `np`.
     # Some functions support `nan_policy='omit'` natively, so setting this to
     # `True` prevents the error from being raised.
-    if xp is None:
-        xp = array_namespace(a)
-    not_numpy = not is_numpy(xp)
-
-    if policies is None:
-        policies = {'propagate', 'raise', 'omit'}
+    policies = {"propagate", "raise", "omit"}
     if nan_policy not in policies:
-        raise ValueError(f"nan_policy must be one of {set(policies)}.")
+        msg = f"nan_policy must be one of {policies}."
+        raise ValueError(msg)
 
     if xp_size(a) == 0:
-        contains_nan = False
-    elif xp.isdtype(a.dtype, "real floating"):
+        return False
+
+    if xp is None:
+        xp = array_namespace(a)
+
+    if xp.isdtype(a.dtype, "real floating"):
         # Faster and less memory-intensive than xp.any(xp.isnan(a)), and unlike other
         # reductions, `max`/`min` won't return NaN unless there is a NaN in the data.
         contains_nan = xp.isnan(xp.max(a))
@@ -958,16 +923,27 @@ def _contains_nan(a, nan_policy='propagate', policies=None, *,
                 break
     else:
         # Only `object` and `inexact` arrays can have NaNs
-        contains_nan = False
-
-    if contains_nan and nan_policy == 'raise':
-        raise ValueError("The input contains nan values")
+        return False
 
-    if not xp_omit_okay and not_numpy and contains_nan and nan_policy=='omit':
-        message = "`nan_policy='omit' is incompatible with non-NumPy arrays."
-        raise ValueError(message)
+    # The implicit call to bool(contains_nan) must happen after testing
+    # nan_policy to prevent lazy and device-bound xps from raising in the
+    # default policy='propagate' case.
+    if nan_policy == 'raise':
+        if is_lazy_array(a):
+            msg = "nan_policy='raise' is not supported for lazy arrays."
+            raise TypeError(msg)
+        if contains_nan:
+            msg = "The input contains nan values"
+            raise ValueError(msg)
+    elif nan_policy == 'omit' and not xp_omit_okay and not is_numpy(xp):
+        if is_lazy_array(a):
+            msg = "nan_policy='omit' is not supported for lazy arrays."
+            raise TypeError(msg)
+        if contains_nan:
+            msg = "nan_policy='omit' is incompatible with non-NumPy arrays."
+            raise ValueError(msg)
 
-    return contains_nan, nan_policy
+    return contains_nan
 
 
 def _rename_parameter(old_name, new_name, dep_version=None):
@@ -1034,17 +1010,16 @@ def _rng_spawn(rng, n_children):
     return child_rngs
 
 
-def _get_nan(*data, xp=None):
+def _get_nan(*data, shape=(), xp=None):
     xp = array_namespace(*data) if xp is None else xp
     # Get NaN of appropriate dtype for data
-    data = [xp.asarray(item) for item in data]
-    try:
-        min_float = getattr(xp, 'float16', xp.float32)
-        dtype = xp.result_type(*data, min_float)  # must be at least a float
-    except DTypePromotionError:
-        # fallback to float64
-        dtype = xp.float64
-    return xp.asarray(xp.nan, dtype=dtype)[()]
+    dtype = xp_result_type(*data, force_floating=True, xp=xp)
+    device = xp_result_device(*data)
+    res = xp.full(shape, xp.nan, dtype=dtype, device=device)
+    if not shape:
+        res = res[()]
+    # whenever mdhaber/marray#89 is resolved, could just return `res`
+    return res.data if is_marray(xp) else res
 
 
 def normalize_axis_index(axis, ndim):
@@ -1177,3 +1152,125 @@ def _dict_formatter(d, n=0, mplus=1, sorter=None):
                              formatter={'float_kind': _float_formatter_10}):
             s = str(d)
     return s
+
+
+_batch_note = """
+The documentation is written assuming array arguments are of specified
+"core" shapes. However, array argument(s) of this function may have additional
+"batch" dimensions prepended to the core shape. In this case, the array is treated
+as a batch of lower-dimensional slices; see :ref:`linalg_batch` for details.
+"""
+
+
+def _apply_over_batch(*argdefs):
+    """
+    Factory for decorator that applies a function over batched arguments.
+
+    Array arguments may have any number of core dimensions (typically 0,
+    1, or 2) and any broadcastable batch shapes. There may be any
+    number of array outputs of any number of dimensions. Assumptions
+    right now - which are satisfied by all functions of interest in `linalg` -
+    are that all array inputs are consecutive keyword or positional arguments,
+    and that the wrapped function returns either a single array or a tuple of
+    arrays. It's only as general as it needs to be right now - it can be extended.
+
+    Parameters
+    ----------
+    *argdefs : tuple of (str, int)
+        Definitions of array arguments: the keyword name of the argument, and
+        the number of core dimensions.
+
+    Example:
+    --------
+    `linalg.eig` accepts two matrices as the first two arguments `a` and `b`, where
+    `b` is optional, and returns one array or a tuple of arrays, depending on the
+    values of other positional or keyword arguments. To generate a wrapper that applies
+    the function over batches of `a` and optionally `b` :
+
+    >>> _apply_over_batch(('a', 2), ('b', 2))
+    """
+    names, ndims = list(zip(*argdefs))
+    n_arrays = len(names)
+
+    def decorator(f):
+        @functools.wraps(f)
+        def wrapper(*args, **kwargs):
+            args = list(args)
+
+            # Ensure all arrays in `arrays`, other arguments in `other_args`/`kwargs`
+            arrays, other_args = args[:n_arrays], args[n_arrays:]
+            for i, name in enumerate(names):
+                if name in kwargs:
+                    if i + 1 <= len(args):
+                        raise ValueError(f'{f.__name__}() got multiple values '
+                                         f'for argument `{name}`.')
+                    else:
+                        arrays.append(kwargs.pop(name))
+
+            # Determine core and batch shapes
+            batch_shapes = []
+            core_shapes = []
+            for i, (array, ndim) in enumerate(zip(arrays, ndims)):
+                array = None if array is None else np.asarray(array)
+                shape = () if array is None else array.shape
+
+                if ndim == "1|2":  # special case for `solve`, etc.
+                    ndim = 2 if array.ndim >= 2 else 1
+
+                arrays[i] = array
+                batch_shapes.append(shape[:-ndim] if ndim > 0 else shape)
+                core_shapes.append(shape[-ndim:] if ndim > 0 else ())
+
+            # Early exit if call is not batched
+            if not any(batch_shapes):
+                return f(*arrays, *other_args, **kwargs)
+
+            # Determine broadcasted batch shape
+            batch_shape = np.broadcast_shapes(*batch_shapes)  # Gives OK error message
+
+            # Broadcast arrays to appropriate shape
+            for i, (array, core_shape) in enumerate(zip(arrays, core_shapes)):
+                if array is None:
+                    continue
+                arrays[i] = np.broadcast_to(array, batch_shape + core_shape)
+
+            # Main loop
+            results = []
+            for index in np.ndindex(batch_shape):
+                result = f(*(array[index] for array in arrays), *other_args, **kwargs)
+                # Assume `result` is either a tuple or single array. This is easily
+                # generalized by allowing the contributor to pass an `unpack_result`
+                # callable to the decorator factory.
+                result = (result,) if not isinstance(result, tuple) else result
+                results.append(result)
+            results = list(zip(*results))
+
+            # Reshape results
+            for i, result in enumerate(results):
+                result = np.stack(result)
+                core_shape = result.shape[1:]
+                results[i] = np.reshape(result, batch_shape + core_shape)
+
+            # Assume `result` should be a single array if there is only one element or
+            # a `tuple` otherwise. This is easily generalized by allowing the
+            # contributor to pass an `pack_result` callable to the decorator factory.
+            return results[0] if len(results) == 1 else results
+
+        doc = FunctionDoc(wrapper)
+        doc['Extended Summary'].append(_batch_note.rstrip())
+        wrapper.__doc__ = str(doc).split("\n", 1)[1]  # remove signature
+
+        return wrapper
+    return decorator
+
+
+def np_vecdot(x1, x2, /, *, axis=-1):
+    # `np.vecdot` has advantages (e.g. see gh-22462), so let's use it when
+    # available. As functions are translated to Array API, `np_vecdot` can be
+    # replaced with `xp.vecdot`.
+    if np.__version__ > "2.0":
+        return np.vecdot(x1, x2, axis=axis)
+    else:
+        # of course there are other fancy ways of doing this (e.g. `einsum`)
+        # but let's keep it simple since it's temporary
+        return np.sum(x1 * x2, axis=axis)

scipy/_lib/array_api_compat/__init__.py

@@ -1,9 +1,9 @@
 """
 NumPy Array API compatibility library
 
-This is a small wrapper around NumPy and CuPy that is compatible with the
-Array API standard https://data-apis.org/array-api/latest/. See also NEP 47
-https://numpy.org/neps/nep-0047-array-api-standard.html.
+This is a small wrapper around NumPy, CuPy, JAX, sparse and others that are
+compatible with the Array API standard https://data-apis.org/array-api/latest/.
+See also NEP 47 https://numpy.org/neps/nep-0047-array-api-standard.html.
 
 Unlike array_api_strict, this is not a strict minimal implementation of the
 Array API, but rather just an extension of the main NumPy namespace with
@@ -17,6 +17,6 @@ to ensure they are not using functionality outside of the standard, but prefer
 this implementation for the default when working with NumPy arrays.
 
 """
-__version__ = '1.9.1'
+__version__ = '1.12.0'
 
 from .common import *  # noqa: F401, F403

scipy/_lib/array_api_compat/_internal.py

@@ -2,10 +2,16 @@
 Internal helpers
 """
 
+from collections.abc import Callable
 from functools import wraps
 from inspect import signature
+from types import ModuleType
+from typing import TypeVar
 
-def get_xp(xp):
+_T = TypeVar("_T")
+
+
+def get_xp(xp: ModuleType) -> Callable[[Callable[..., _T]], Callable[..., _T]]:
     """
     Decorator to automatically replace xp with the corresponding array module.
 
@@ -22,14 +28,14 @@ def get_xp(xp):
 
     """
 
-    def inner(f):
+    def inner(f: Callable[..., _T], /) -> Callable[..., _T]:
         @wraps(f)
-        def wrapped_f(*args, **kwargs):
+        def wrapped_f(*args: object, **kwargs: object) -> object:
             return f(*args, xp=xp, **kwargs)
 
         sig = signature(f)
         new_sig = sig.replace(
-            parameters=[sig.parameters[i] for i in sig.parameters if i != "xp"]
+            parameters=[par for i, par in sig.parameters.items() if i != "xp"]
         )
 
         if wrapped_f.__doc__ is None:
@@ -40,7 +46,14 @@ See the corresponding documentation in NumPy/CuPy and/or the array API
 specification for more details.
 
 """
-        wrapped_f.__signature__ = new_sig
-        return wrapped_f
+        wrapped_f.__signature__ = new_sig  # pyright: ignore[reportAttributeAccessIssue]
+        return wrapped_f  # pyright: ignore[reportReturnType]
 
     return inner
+
+
+__all__ = ["get_xp"]
+
+
+def __dir__() -> list[str]:
+    return __all__

scipy/_lib/array_api_compat/common/__init__.py

@@ -1 +1 @@
-from ._helpers import * # noqa: F403
+from ._helpers import *  # noqa: F403