Add files using upload-large-folder tool

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/core/cversions.py

@@ -0,0 +1,13 @@
+"""Simple script to compute the api hash of the current API.
+
+The API has is defined by numpy_api_order and ufunc_api_order.
+
+"""
+from os.path import dirname
+
+from code_generators.genapi import fullapi_hash
+from code_generators.numpy_api import full_api
+
+if __name__ == '__main__':
+    curdir = dirname(__file__)
+    print(fullapi_hash(full_api))

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/core/defchararray.py

@@ -0,0 +1,2914 @@
+"""
+This module contains a set of functions for vectorized string
+operations and methods.
+
+.. note::
+   The `chararray` class exists for backwards compatibility with
+   Numarray, it is not recommended for new development. Starting from numpy
+   1.4, if one needs arrays of strings, it is recommended to use arrays of
+   `dtype` `object_`, `bytes_` or `str_`, and use the free functions
+   in the `numpy.char` module for fast vectorized string operations.
+
+Some methods will only be available if the corresponding string method is
+available in your version of Python.
+
+The preferred alias for `defchararray` is `numpy.char`.
+
+"""
+import functools
+
+from .._utils import set_module
+from .numerictypes import (
+    bytes_, str_, integer, int_, object_, bool_, character)
+from .numeric import ndarray, compare_chararrays
+from .numeric import array as narray
+from numpy.core.multiarray import _vec_string
+from numpy.core import overrides
+from numpy.compat import asbytes
+import numpy
+
+__all__ = [
+    'equal', 'not_equal', 'greater_equal', 'less_equal',
+    'greater', 'less', 'str_len', 'add', 'multiply', 'mod', 'capitalize',
+    'center', 'count', 'decode', 'encode', 'endswith', 'expandtabs',
+    'find', 'index', 'isalnum', 'isalpha', 'isdigit', 'islower', 'isspace',
+    'istitle', 'isupper', 'join', 'ljust', 'lower', 'lstrip', 'partition',
+    'replace', 'rfind', 'rindex', 'rjust', 'rpartition', 'rsplit',
+    'rstrip', 'split', 'splitlines', 'startswith', 'strip', 'swapcase',
+    'title', 'translate', 'upper', 'zfill', 'isnumeric', 'isdecimal',
+    'array', 'asarray'
+    ]
+
+
+_globalvar = 0
+
+array_function_dispatch = functools.partial(
+    overrides.array_function_dispatch, module='numpy.char')
+
+
+def _is_unicode(arr):
+    """Returns True if arr is a string or a string array with a dtype that
+    represents a unicode string, otherwise returns False.
+
+    """
+    if (isinstance(arr, str) or
+            issubclass(numpy.asarray(arr).dtype.type, str)):
+        return True
+    return False
+
+
+def _to_bytes_or_str_array(result, output_dtype_like=None):
+    """
+    Helper function to cast a result back into an array
+    with the appropriate dtype if an object array must be used
+    as an intermediary.
+    """
+    ret = numpy.asarray(result.tolist())
+    dtype = getattr(output_dtype_like, 'dtype', None)
+    if dtype is not None:
+        return ret.astype(type(dtype)(_get_num_chars(ret)), copy=False)
+    return ret
+
+
+def _clean_args(*args):
+    """
+    Helper function for delegating arguments to Python string
+    functions.
+
+    Many of the Python string operations that have optional arguments
+    do not use 'None' to indicate a default value.  In these cases,
+    we need to remove all None arguments, and those following them.
+    """
+    newargs = []
+    for chk in args:
+        if chk is None:
+            break
+        newargs.append(chk)
+    return newargs
+
+def _get_num_chars(a):
+    """
+    Helper function that returns the number of characters per field in
+    a string or unicode array.  This is to abstract out the fact that
+    for a unicode array this is itemsize / 4.
+    """
+    if issubclass(a.dtype.type, str_):
+        return a.itemsize // 4
+    return a.itemsize
+
+
+def _binary_op_dispatcher(x1, x2):
+    return (x1, x2)
+
+
+@array_function_dispatch(_binary_op_dispatcher)
+def equal(x1, x2):
+    """
+    Return (x1 == x2) element-wise.
+
+    Unlike `numpy.equal`, this comparison is performed by first
+    stripping whitespace characters from the end of the string.  This
+    behavior is provided for backward-compatibility with numarray.
+
+    Parameters
+    ----------
+    x1, x2 : array_like of str or unicode
+        Input arrays of the same shape.
+
+    Returns
+    -------
+    out : ndarray
+        Output array of bools.
+
+    See Also
+    --------
+    not_equal, greater_equal, less_equal, greater, less
+    """
+    return compare_chararrays(x1, x2, '==', True)
+
+
+@array_function_dispatch(_binary_op_dispatcher)
+def not_equal(x1, x2):
+    """
+    Return (x1 != x2) element-wise.
+
+    Unlike `numpy.not_equal`, this comparison is performed by first
+    stripping whitespace characters from the end of the string.  This
+    behavior is provided for backward-compatibility with numarray.
+
+    Parameters
+    ----------
+    x1, x2 : array_like of str or unicode
+        Input arrays of the same shape.
+
+    Returns
+    -------
+    out : ndarray
+        Output array of bools.
+
+    See Also
+    --------
+    equal, greater_equal, less_equal, greater, less
+    """
+    return compare_chararrays(x1, x2, '!=', True)
+
+
+@array_function_dispatch(_binary_op_dispatcher)
+def greater_equal(x1, x2):
+    """
+    Return (x1 >= x2) element-wise.
+
+    Unlike `numpy.greater_equal`, this comparison is performed by
+    first stripping whitespace characters from the end of the string.
+    This behavior is provided for backward-compatibility with
+    numarray.
+
+    Parameters
+    ----------
+    x1, x2 : array_like of str or unicode
+        Input arrays of the same shape.
+
+    Returns
+    -------
+    out : ndarray
+        Output array of bools.
+
+    See Also
+    --------
+    equal, not_equal, less_equal, greater, less
+    """
+    return compare_chararrays(x1, x2, '>=', True)
+
+
+@array_function_dispatch(_binary_op_dispatcher)
+def less_equal(x1, x2):
+    """
+    Return (x1 <= x2) element-wise.
+
+    Unlike `numpy.less_equal`, this comparison is performed by first
+    stripping whitespace characters from the end of the string.  This
+    behavior is provided for backward-compatibility with numarray.
+
+    Parameters
+    ----------
+    x1, x2 : array_like of str or unicode
+        Input arrays of the same shape.
+
+    Returns
+    -------
+    out : ndarray
+        Output array of bools.
+
+    See Also
+    --------
+    equal, not_equal, greater_equal, greater, less
+    """
+    return compare_chararrays(x1, x2, '<=', True)
+
+
+@array_function_dispatch(_binary_op_dispatcher)
+def greater(x1, x2):
+    """
+    Return (x1 > x2) element-wise.
+
+    Unlike `numpy.greater`, this comparison is performed by first
+    stripping whitespace characters from the end of the string.  This
+    behavior is provided for backward-compatibility with numarray.
+
+    Parameters
+    ----------
+    x1, x2 : array_like of str or unicode
+        Input arrays of the same shape.
+
+    Returns
+    -------
+    out : ndarray
+        Output array of bools.
+
+    See Also
+    --------
+    equal, not_equal, greater_equal, less_equal, less
+    """
+    return compare_chararrays(x1, x2, '>', True)
+
+
+@array_function_dispatch(_binary_op_dispatcher)
+def less(x1, x2):
+    """
+    Return (x1 < x2) element-wise.
+
+    Unlike `numpy.greater`, this comparison is performed by first
+    stripping whitespace characters from the end of the string.  This
+    behavior is provided for backward-compatibility with numarray.
+
+    Parameters
+    ----------
+    x1, x2 : array_like of str or unicode
+        Input arrays of the same shape.
+
+    Returns
+    -------
+    out : ndarray
+        Output array of bools.
+
+    See Also
+    --------
+    equal, not_equal, greater_equal, less_equal, greater
+    """
+    return compare_chararrays(x1, x2, '<', True)
+
+
+def _unary_op_dispatcher(a):
+    return (a,)
+
+
+@array_function_dispatch(_unary_op_dispatcher)
+def str_len(a):
+    """
+    Return len(a) element-wise.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    Returns
+    -------
+    out : ndarray
+        Output array of integers
+
+    See Also
+    --------
+    len
+
+    Examples
+    --------
+    >>> a = np.array(['Grace Hopper Conference', 'Open Source Day'])
+    >>> np.char.str_len(a)
+    array([23, 15])
+    >>> a = np.array([u'\u0420', u'\u043e'])
+    >>> np.char.str_len(a)
+    array([1, 1])
+    >>> a = np.array([['hello', 'world'], [u'\u0420', u'\u043e']])
+    >>> np.char.str_len(a)
+    array([[5, 5], [1, 1]])
+    """
+    # Note: __len__, etc. currently return ints, which are not C-integers.
+    # Generally intp would be expected for lengths, although int is sufficient
+    # due to the dtype itemsize limitation.
+    return _vec_string(a, int_, '__len__')
+
+
+@array_function_dispatch(_binary_op_dispatcher)
+def add(x1, x2):
+    """
+    Return element-wise string concatenation for two arrays of str or unicode.
+
+    Arrays `x1` and `x2` must have the same shape.
+
+    Parameters
+    ----------
+    x1 : array_like of str or unicode
+        Input array.
+    x2 : array_like of str or unicode
+        Input array.
+
+    Returns
+    -------
+    add : ndarray
+        Output array of `bytes_` or `str_`, depending on input types
+        of the same shape as `x1` and `x2`.
+
+    """
+    arr1 = numpy.asarray(x1)
+    arr2 = numpy.asarray(x2)
+    out_size = _get_num_chars(arr1) + _get_num_chars(arr2)
+
+    if type(arr1.dtype) != type(arr2.dtype):
+        # Enforce this for now.  The solution to it will be implement add
+        # as a ufunc.  It never worked right on Python 3: bytes + unicode gave
+        # nonsense unicode + bytes errored, and unicode + object used the
+        # object dtype itemsize as num chars (worked on short strings).
+        # bytes + void worked but promoting void->bytes is dubious also.
+        raise TypeError(
+            "np.char.add() requires both arrays of the same dtype kind, but "
+            f"got dtypes: '{arr1.dtype}' and '{arr2.dtype}' (the few cases "
+            "where this used to work often lead to incorrect results).")
+
+    return _vec_string(arr1, type(arr1.dtype)(out_size), '__add__', (arr2,))
+
+def _multiply_dispatcher(a, i):
+    return (a,)
+
+
+@array_function_dispatch(_multiply_dispatcher)
+def multiply(a, i):
+    """
+    Return (a * i), that is string multiple concatenation,
+    element-wise.
+
+    Values in `i` of less than 0 are treated as 0 (which yields an
+    empty string).
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    i : array_like of ints
+
+    Returns
+    -------
+    out : ndarray
+        Output array of str or unicode, depending on input types
+    
+    Examples
+    --------
+    >>> a = np.array(["a", "b", "c"])
+    >>> np.char.multiply(x, 3)
+    array(['aaa', 'bbb', 'ccc'], dtype='<U3')
+    >>> i = np.array([1, 2, 3])
+    >>> np.char.multiply(a, i)
+    array(['a', 'bb', 'ccc'], dtype='<U3')
+    >>> np.char.multiply(np.array(['a']), i)
+    array(['a', 'aa', 'aaa'], dtype='<U3')
+    >>> a = np.array(['a', 'b', 'c', 'd', 'e', 'f']).reshape((2, 3))
+    >>> np.char.multiply(a, 3)
+    array([['aaa', 'bbb', 'ccc'],
+           ['ddd', 'eee', 'fff']], dtype='<U3')
+    >>> np.char.multiply(a, i)
+    array([['a', 'bb', 'ccc'],
+           ['d', 'ee', 'fff']], dtype='<U3')
+    """
+    a_arr = numpy.asarray(a)
+    i_arr = numpy.asarray(i)
+    if not issubclass(i_arr.dtype.type, integer):
+        raise ValueError("Can only multiply by integers")
+    out_size = _get_num_chars(a_arr) * max(int(i_arr.max()), 0)
+    return _vec_string(
+        a_arr, type(a_arr.dtype)(out_size), '__mul__', (i_arr,))
+
+
+def _mod_dispatcher(a, values):
+    return (a, values)
+
+
+@array_function_dispatch(_mod_dispatcher)
+def mod(a, values):
+    """
+    Return (a % i), that is pre-Python 2.6 string formatting
+    (interpolation), element-wise for a pair of array_likes of str
+    or unicode.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    values : array_like of values
+       These values will be element-wise interpolated into the string.
+
+    Returns
+    -------
+    out : ndarray
+        Output array of str or unicode, depending on input types
+
+    See Also
+    --------
+    str.__mod__
+
+    """
+    return _to_bytes_or_str_array(
+        _vec_string(a, object_, '__mod__', (values,)), a)
+
+
+@array_function_dispatch(_unary_op_dispatcher)
+def capitalize(a):
+    """
+    Return a copy of `a` with only the first character of each element
+    capitalized.
+
+    Calls `str.capitalize` element-wise.
+
+    For 8-bit strings, this method is locale-dependent.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+        Input array of strings to capitalize.
+
+    Returns
+    -------
+    out : ndarray
+        Output array of str or unicode, depending on input
+        types
+
+    See Also
+    --------
+    str.capitalize
+
+    Examples
+    --------
+    >>> c = np.array(['a1b2','1b2a','b2a1','2a1b'],'S4'); c
+    array(['a1b2', '1b2a', 'b2a1', '2a1b'],
+        dtype='|S4')
+    >>> np.char.capitalize(c)
+    array(['A1b2', '1b2a', 'B2a1', '2a1b'],
+        dtype='|S4')
+
+    """
+    a_arr = numpy.asarray(a)
+    return _vec_string(a_arr, a_arr.dtype, 'capitalize')
+
+
+def _center_dispatcher(a, width, fillchar=None):
+    return (a,)
+
+
+@array_function_dispatch(_center_dispatcher)
+def center(a, width, fillchar=' '):
+    """
+    Return a copy of `a` with its elements centered in a string of
+    length `width`.
+
+    Calls `str.center` element-wise.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    width : int
+        The length of the resulting strings
+    fillchar : str or unicode, optional
+        The padding character to use (default is space).
+
+    Returns
+    -------
+    out : ndarray
+        Output array of str or unicode, depending on input
+        types
+
+    See Also
+    --------
+    str.center
+    
+    Notes
+    -----
+    This function is intended to work with arrays of strings.  The
+    fill character is not applied to numeric types.
+
+    Examples
+    --------
+    >>> c = np.array(['a1b2','1b2a','b2a1','2a1b']); c
+    array(['a1b2', '1b2a', 'b2a1', '2a1b'], dtype='<U4')
+    >>> np.char.center(c, width=9)
+    array(['   a1b2  ', '   1b2a  ', '   b2a1  ', '   2a1b  '], dtype='<U9')
+    >>> np.char.center(c, width=9, fillchar='*')
+    array(['***a1b2**', '***1b2a**', '***b2a1**', '***2a1b**'], dtype='<U9')
+    >>> np.char.center(c, width=1)
+    array(['a', '1', 'b', '2'], dtype='<U1')
+
+    """
+    a_arr = numpy.asarray(a)
+    width_arr = numpy.asarray(width)
+    size = int(numpy.max(width_arr.flat))
+    if numpy.issubdtype(a_arr.dtype, numpy.bytes_):
+        fillchar = asbytes(fillchar)
+    return _vec_string(
+        a_arr, type(a_arr.dtype)(size), 'center', (width_arr, fillchar))
+
+
+def _count_dispatcher(a, sub, start=None, end=None):
+    return (a,)
+
+
+@array_function_dispatch(_count_dispatcher)
+def count(a, sub, start=0, end=None):
+    """
+    Returns an array with the number of non-overlapping occurrences of
+    substring `sub` in the range [`start`, `end`].
+
+    Calls `str.count` element-wise.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    sub : str or unicode
+       The substring to search for.
+
+    start, end : int, optional
+       Optional arguments `start` and `end` are interpreted as slice
+       notation to specify the range in which to count.
+
+    Returns
+    -------
+    out : ndarray
+        Output array of ints.
+
+    See Also
+    --------
+    str.count
+
+    Examples
+    --------
+    >>> c = np.array(['aAaAaA', '  aA  ', 'abBABba'])
+    >>> c
+    array(['aAaAaA', '  aA  ', 'abBABba'], dtype='<U7')
+    >>> np.char.count(c, 'A')
+    array([3, 1, 1])
+    >>> np.char.count(c, 'aA')
+    array([3, 1, 0])
+    >>> np.char.count(c, 'A', start=1, end=4)
+    array([2, 1, 1])
+    >>> np.char.count(c, 'A', start=1, end=3)
+    array([1, 0, 0])
+
+    """
+    return _vec_string(a, int_, 'count', [sub, start] + _clean_args(end))
+
+
+def _code_dispatcher(a, encoding=None, errors=None):
+    return (a,)
+
+
+@array_function_dispatch(_code_dispatcher)
+def decode(a, encoding=None, errors=None):
+    r"""
+    Calls ``bytes.decode`` element-wise.
+
+    The set of available codecs comes from the Python standard library,
+    and may be extended at runtime.  For more information, see the
+    :mod:`codecs` module.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    encoding : str, optional
+       The name of an encoding
+
+    errors : str, optional
+       Specifies how to handle encoding errors
+
+    Returns
+    -------
+    out : ndarray
+
+    See Also
+    --------
+    :py:meth:`bytes.decode`
+
+    Notes
+    -----
+    The type of the result will depend on the encoding specified.
+
+    Examples
+    --------
+    >>> c = np.array([b'\x81\xc1\x81\xc1\x81\xc1', b'@@\x81\xc1@@',
+    ...               b'\x81\x82\xc2\xc1\xc2\x82\x81'])
+    >>> c
+    array([b'\x81\xc1\x81\xc1\x81\xc1', b'@@\x81\xc1@@',
+    ...    b'\x81\x82\xc2\xc1\xc2\x82\x81'], dtype='|S7')
+    >>> np.char.decode(c, encoding='cp037')
+    array(['aAaAaA', '  aA  ', 'abBABba'], dtype='<U7')
+
+    """
+    return _to_bytes_or_str_array(
+        _vec_string(a, object_, 'decode', _clean_args(encoding, errors)))
+
+
+@array_function_dispatch(_code_dispatcher)
+def encode(a, encoding=None, errors=None):
+    """
+    Calls `str.encode` element-wise.
+
+    The set of available codecs comes from the Python standard library,
+    and may be extended at runtime. For more information, see the codecs
+    module.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    encoding : str, optional
+       The name of an encoding
+
+    errors : str, optional
+       Specifies how to handle encoding errors
+
+    Returns
+    -------
+    out : ndarray
+
+    See Also
+    --------
+    str.encode
+
+    Notes
+    -----
+    The type of the result will depend on the encoding specified.
+
+    """
+    return _to_bytes_or_str_array(
+        _vec_string(a, object_, 'encode', _clean_args(encoding, errors)))
+
+
+def _endswith_dispatcher(a, suffix, start=None, end=None):
+    return (a,)
+
+
+@array_function_dispatch(_endswith_dispatcher)
+def endswith(a, suffix, start=0, end=None):
+    """
+    Returns a boolean array which is `True` where the string element
+    in `a` ends with `suffix`, otherwise `False`.
+
+    Calls `str.endswith` element-wise.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    suffix : str
+
+    start, end : int, optional
+        With optional `start`, test beginning at that position. With
+        optional `end`, stop comparing at that position.
+
+    Returns
+    -------
+    out : ndarray
+        Outputs an array of bools.
+
+    See Also
+    --------
+    str.endswith
+
+    Examples
+    --------
+    >>> s = np.array(['foo', 'bar'])
+    >>> s[0] = 'foo'
+    >>> s[1] = 'bar'
+    >>> s
+    array(['foo', 'bar'], dtype='<U3')
+    >>> np.char.endswith(s, 'ar')
+    array([False,  True])
+    >>> np.char.endswith(s, 'a', start=1, end=2)
+    array([False,  True])
+
+    """
+    return _vec_string(
+        a, bool_, 'endswith', [suffix, start] + _clean_args(end))
+
+
+def _expandtabs_dispatcher(a, tabsize=None):
+    return (a,)
+
+
+@array_function_dispatch(_expandtabs_dispatcher)
+def expandtabs(a, tabsize=8):
+    """
+    Return a copy of each string element where all tab characters are
+    replaced by one or more spaces.
+
+    Calls `str.expandtabs` element-wise.
+
+    Return a copy of each string element where all tab characters are
+    replaced by one or more spaces, depending on the current column
+    and the given `tabsize`. The column number is reset to zero after
+    each newline occurring in the string. This doesn't understand other
+    non-printing characters or escape sequences.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+        Input array
+    tabsize : int, optional
+        Replace tabs with `tabsize` number of spaces.  If not given defaults
+        to 8 spaces.
+
+    Returns
+    -------
+    out : ndarray
+        Output array of str or unicode, depending on input type
+
+    See Also
+    --------
+    str.expandtabs
+
+    """
+    return _to_bytes_or_str_array(
+        _vec_string(a, object_, 'expandtabs', (tabsize,)), a)
+
+
+@array_function_dispatch(_count_dispatcher)
+def find(a, sub, start=0, end=None):
+    """
+    For each element, return the lowest index in the string where
+    substring `sub` is found.
+
+    Calls `str.find` element-wise.
+
+    For each element, return the lowest index in the string where
+    substring `sub` is found, such that `sub` is contained in the
+    range [`start`, `end`].
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    sub : str or unicode
+
+    start, end : int, optional
+        Optional arguments `start` and `end` are interpreted as in
+        slice notation.
+
+    Returns
+    -------
+    out : ndarray or int
+        Output array of ints.  Returns -1 if `sub` is not found.
+
+    See Also
+    --------
+    str.find
+
+    Examples
+    --------
+    >>> a = np.array(["NumPy is a Python library"])
+    >>> np.char.find(a, "Python", start=0, end=None)
+    array([11])
+
+    """
+    return _vec_string(
+        a, int_, 'find', [sub, start] + _clean_args(end))
+
+
+@array_function_dispatch(_count_dispatcher)
+def index(a, sub, start=0, end=None):
+    """
+    Like `find`, but raises `ValueError` when the substring is not found.
+
+    Calls `str.index` element-wise.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    sub : str or unicode
+
+    start, end : int, optional
+
+    Returns
+    -------
+    out : ndarray
+        Output array of ints.  Returns -1 if `sub` is not found.
+
+    See Also
+    --------
+    find, str.find
+
+    Examples
+    --------
+    >>> a = np.array(["Computer Science"])
+    >>> np.char.index(a, "Science", start=0, end=None)
+    array([9])
+
+    """
+    return _vec_string(
+        a, int_, 'index', [sub, start] + _clean_args(end))
+
+
+@array_function_dispatch(_unary_op_dispatcher)
+def isalnum(a):
+    """
+    Returns true for each element if all characters in the string are
+    alphanumeric and there is at least one character, false otherwise.
+
+    Calls `str.isalnum` element-wise.
+
+    For 8-bit strings, this method is locale-dependent.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    Returns
+    -------
+    out : ndarray
+        Output array of str or unicode, depending on input type
+
+    See Also
+    --------
+    str.isalnum
+    """
+    return _vec_string(a, bool_, 'isalnum')
+
+
+@array_function_dispatch(_unary_op_dispatcher)
+def isalpha(a):
+    """
+    Returns true for each element if all characters in the string are
+    alphabetic and there is at least one character, false otherwise.
+
+    Calls `str.isalpha` element-wise.
+
+    For 8-bit strings, this method is locale-dependent.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    Returns
+    -------
+    out : ndarray
+        Output array of bools
+
+    See Also
+    --------
+    str.isalpha
+    """
+    return _vec_string(a, bool_, 'isalpha')
+
+
+@array_function_dispatch(_unary_op_dispatcher)
+def isdigit(a):
+    """
+    Returns true for each element if all characters in the string are
+    digits and there is at least one character, false otherwise.
+
+    Calls `str.isdigit` element-wise.
+
+    For 8-bit strings, this method is locale-dependent.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    Returns
+    -------
+    out : ndarray
+        Output array of bools
+
+    See Also
+    --------
+    str.isdigit
+
+    Examples
+    --------
+    >>> a = np.array(['a', 'b', '0'])
+    >>> np.char.isdigit(a)
+    array([False, False,  True])
+    >>> a = np.array([['a', 'b', '0'], ['c', '1', '2']])
+    >>> np.char.isdigit(a)
+    array([[False, False,  True], [False,  True,  True]])
+    """
+    return _vec_string(a, bool_, 'isdigit')
+
+
+@array_function_dispatch(_unary_op_dispatcher)
+def islower(a):
+    """
+    Returns true for each element if all cased characters in the
+    string are lowercase and there is at least one cased character,
+    false otherwise.
+
+    Calls `str.islower` element-wise.
+
+    For 8-bit strings, this method is locale-dependent.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    Returns
+    -------
+    out : ndarray
+        Output array of bools
+
+    See Also
+    --------
+    str.islower
+    """
+    return _vec_string(a, bool_, 'islower')
+
+
+@array_function_dispatch(_unary_op_dispatcher)
+def isspace(a):
+    """
+    Returns true for each element if there are only whitespace
+    characters in the string and there is at least one character,
+    false otherwise.
+
+    Calls `str.isspace` element-wise.
+
+    For 8-bit strings, this method is locale-dependent.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    Returns
+    -------
+    out : ndarray
+        Output array of bools
+
+    See Also
+    --------
+    str.isspace
+    """
+    return _vec_string(a, bool_, 'isspace')
+
+
+@array_function_dispatch(_unary_op_dispatcher)
+def istitle(a):
+    """
+    Returns true for each element if the element is a titlecased
+    string and there is at least one character, false otherwise.
+
+    Call `str.istitle` element-wise.
+
+    For 8-bit strings, this method is locale-dependent.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    Returns
+    -------
+    out : ndarray
+        Output array of bools
+
+    See Also
+    --------
+    str.istitle
+    """
+    return _vec_string(a, bool_, 'istitle')
+
+
+@array_function_dispatch(_unary_op_dispatcher)
+def isupper(a):
+    """
+    Return true for each element if all cased characters in the
+    string are uppercase and there is at least one character, false
+    otherwise.
+
+    Call `str.isupper` element-wise.
+
+    For 8-bit strings, this method is locale-dependent.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    Returns
+    -------
+    out : ndarray
+        Output array of bools
+
+    See Also
+    --------
+    str.isupper
+
+    Examples
+    --------
+    >>> str = "GHC"
+    >>> np.char.isupper(str)
+    array(True)     
+    >>> a = np.array(["hello", "HELLO", "Hello"])
+    >>> np.char.isupper(a)
+    array([False,  True, False]) 
+
+    """
+    return _vec_string(a, bool_, 'isupper')
+
+
+def _join_dispatcher(sep, seq):
+    return (sep, seq)
+
+
+@array_function_dispatch(_join_dispatcher)
+def join(sep, seq):
+    """
+    Return a string which is the concatenation of the strings in the
+    sequence `seq`.
+
+    Calls `str.join` element-wise.
+
+    Parameters
+    ----------
+    sep : array_like of str or unicode
+    seq : array_like of str or unicode
+
+    Returns
+    -------
+    out : ndarray
+        Output array of str or unicode, depending on input types
+
+    See Also
+    --------
+    str.join
+
+    Examples
+    --------
+    >>> np.char.join('-', 'osd')
+    array('o-s-d', dtype='<U5')
+
+    >>> np.char.join(['-', '.'], ['ghc', 'osd'])
+    array(['g-h-c', 'o.s.d'], dtype='<U5')
+
+    """
+    return _to_bytes_or_str_array(
+        _vec_string(sep, object_, 'join', (seq,)), seq)
+
+
+
+def _just_dispatcher(a, width, fillchar=None):
+    return (a,)
+
+
+@array_function_dispatch(_just_dispatcher)
+def ljust(a, width, fillchar=' '):
+    """
+    Return an array with the elements of `a` left-justified in a
+    string of length `width`.
+
+    Calls `str.ljust` element-wise.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    width : int
+        The length of the resulting strings
+    fillchar : str or unicode, optional
+        The character to use for padding
+
+    Returns
+    -------
+    out : ndarray
+        Output array of str or unicode, depending on input type
+
+    See Also
+    --------
+    str.ljust
+
+    """
+    a_arr = numpy.asarray(a)
+    width_arr = numpy.asarray(width)
+    size = int(numpy.max(width_arr.flat))
+    if numpy.issubdtype(a_arr.dtype, numpy.bytes_):
+        fillchar = asbytes(fillchar)
+    return _vec_string(
+        a_arr, type(a_arr.dtype)(size), 'ljust', (width_arr, fillchar))
+
+
+@array_function_dispatch(_unary_op_dispatcher)
+def lower(a):
+    """
+    Return an array with the elements converted to lowercase.
+
+    Call `str.lower` element-wise.
+
+    For 8-bit strings, this method is locale-dependent.
+
+    Parameters
+    ----------
+    a : array_like, {str, unicode}
+        Input array.
+
+    Returns
+    -------
+    out : ndarray, {str, unicode}
+        Output array of str or unicode, depending on input type
+
+    See Also
+    --------
+    str.lower
+
+    Examples
+    --------
+    >>> c = np.array(['A1B C', '1BCA', 'BCA1']); c
+    array(['A1B C', '1BCA', 'BCA1'], dtype='<U5')
+    >>> np.char.lower(c)
+    array(['a1b c', '1bca', 'bca1'], dtype='<U5')
+
+    """
+    a_arr = numpy.asarray(a)
+    return _vec_string(a_arr, a_arr.dtype, 'lower')
+
+
+def _strip_dispatcher(a, chars=None):
+    return (a,)
+
+
+@array_function_dispatch(_strip_dispatcher)
+def lstrip(a, chars=None):
+    """
+    For each element in `a`, return a copy with the leading characters
+    removed.
+
+    Calls `str.lstrip` element-wise.
+
+    Parameters
+    ----------
+    a : array-like, {str, unicode}
+        Input array.
+
+    chars : {str, unicode}, optional
+        The `chars` argument is a string specifying the set of
+        characters to be removed. If omitted or None, the `chars`
+        argument defaults to removing whitespace. The `chars` argument
+        is not a prefix; rather, all combinations of its values are
+        stripped.
+
+    Returns
+    -------
+    out : ndarray, {str, unicode}
+        Output array of str or unicode, depending on input type
+
+    See Also
+    --------
+    str.lstrip
+
+    Examples
+    --------
+    >>> c = np.array(['aAaAaA', '  aA  ', 'abBABba'])
+    >>> c
+    array(['aAaAaA', '  aA  ', 'abBABba'], dtype='<U7')
+
+    The 'a' variable is unstripped from c[1] because whitespace leading.
+
+    >>> np.char.lstrip(c, 'a')
+    array(['AaAaA', '  aA  ', 'bBABba'], dtype='<U7')
+
+
+    >>> np.char.lstrip(c, 'A') # leaves c unchanged
+    array(['aAaAaA', '  aA  ', 'abBABba'], dtype='<U7')
+    >>> (np.char.lstrip(c, ' ') == np.char.lstrip(c, '')).all()
+    ... # XXX: is this a regression? This used to return True
+    ... # np.char.lstrip(c,'') does not modify c at all.
+    False
+    >>> (np.char.lstrip(c, ' ') == np.char.lstrip(c, None)).all()
+    True
+
+    """
+    a_arr = numpy.asarray(a)
+    return _vec_string(a_arr, a_arr.dtype, 'lstrip', (chars,))
+
+
+def _partition_dispatcher(a, sep):
+    return (a,)
+
+
+@array_function_dispatch(_partition_dispatcher)
+def partition(a, sep):
+    """
+    Partition each element in `a` around `sep`.
+
+    Calls `str.partition` element-wise.
+
+    For each element in `a`, split the element as the first
+    occurrence of `sep`, and return 3 strings containing the part
+    before the separator, the separator itself, and the part after
+    the separator. If the separator is not found, return 3 strings
+    containing the string itself, followed by two empty strings.
+
+    Parameters
+    ----------
+    a : array_like, {str, unicode}
+        Input array
+    sep : {str, unicode}
+        Separator to split each string element in `a`.
+
+    Returns
+    -------
+    out : ndarray, {str, unicode}
+        Output array of str or unicode, depending on input type.
+        The output array will have an extra dimension with 3
+        elements per input element.
+
+    See Also
+    --------
+    str.partition
+
+    """
+    return _to_bytes_or_str_array(
+        _vec_string(a, object_, 'partition', (sep,)), a)
+
+
+def _replace_dispatcher(a, old, new, count=None):
+    return (a,)
+
+
+@array_function_dispatch(_replace_dispatcher)
+def replace(a, old, new, count=None):
+    """
+    For each element in `a`, return a copy of the string with all
+    occurrences of substring `old` replaced by `new`.
+
+    Calls `str.replace` element-wise.
+
+    Parameters
+    ----------
+    a : array-like of str or unicode
+
+    old, new : str or unicode
+
+    count : int, optional
+        If the optional argument `count` is given, only the first
+        `count` occurrences are replaced.
+
+    Returns
+    -------
+    out : ndarray
+        Output array of str or unicode, depending on input type
+
+    See Also
+    --------
+    str.replace
+    
+    Examples
+    --------
+    >>> a = np.array(["That is a mango", "Monkeys eat mangos"])
+    >>> np.char.replace(a, 'mango', 'banana')
+    array(['That is a banana', 'Monkeys eat bananas'], dtype='<U19')
+
+    >>> a = np.array(["The dish is fresh", "This is it"])
+    >>> np.char.replace(a, 'is', 'was')
+    array(['The dwash was fresh', 'Thwas was it'], dtype='<U19')
+    """
+    return _to_bytes_or_str_array(
+        _vec_string(a, object_, 'replace', [old, new] + _clean_args(count)), a)
+
+
+@array_function_dispatch(_count_dispatcher)
+def rfind(a, sub, start=0, end=None):
+    """
+    For each element in `a`, return the highest index in the string
+    where substring `sub` is found, such that `sub` is contained
+    within [`start`, `end`].
+
+    Calls `str.rfind` element-wise.
+
+    Parameters
+    ----------
+    a : array-like of str or unicode
+
+    sub : str or unicode
+
+    start, end : int, optional
+        Optional arguments `start` and `end` are interpreted as in
+        slice notation.
+
+    Returns
+    -------
+    out : ndarray
+       Output array of ints.  Return -1 on failure.
+
+    See Also
+    --------
+    str.rfind
+
+    """
+    return _vec_string(
+        a, int_, 'rfind', [sub, start] + _clean_args(end))
+
+
+@array_function_dispatch(_count_dispatcher)
+def rindex(a, sub, start=0, end=None):
+    """
+    Like `rfind`, but raises `ValueError` when the substring `sub` is
+    not found.
+
+    Calls `str.rindex` element-wise.
+
+    Parameters
+    ----------
+    a : array-like of str or unicode
+
+    sub : str or unicode
+
+    start, end : int, optional
+
+    Returns
+    -------
+    out : ndarray
+       Output array of ints.
+
+    See Also
+    --------
+    rfind, str.rindex
+
+    """
+    return _vec_string(
+        a, int_, 'rindex', [sub, start] + _clean_args(end))
+
+
+@array_function_dispatch(_just_dispatcher)
+def rjust(a, width, fillchar=' '):
+    """
+    Return an array with the elements of `a` right-justified in a
+    string of length `width`.
+
+    Calls `str.rjust` element-wise.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    width : int
+        The length of the resulting strings
+    fillchar : str or unicode, optional
+        The character to use for padding
+
+    Returns
+    -------
+    out : ndarray
+        Output array of str or unicode, depending on input type
+
+    See Also
+    --------
+    str.rjust
+
+    """
+    a_arr = numpy.asarray(a)
+    width_arr = numpy.asarray(width)
+    size = int(numpy.max(width_arr.flat))
+    if numpy.issubdtype(a_arr.dtype, numpy.bytes_):
+        fillchar = asbytes(fillchar)
+    return _vec_string(
+        a_arr, type(a_arr.dtype)(size), 'rjust', (width_arr, fillchar))
+
+
+@array_function_dispatch(_partition_dispatcher)
+def rpartition(a, sep):
+    """
+    Partition (split) each element around the right-most separator.
+
+    Calls `str.rpartition` element-wise.
+
+    For each element in `a`, split the element as the last
+    occurrence of `sep`, and return 3 strings containing the part
+    before the separator, the separator itself, and the part after
+    the separator. If the separator is not found, return 3 strings
+    containing the string itself, followed by two empty strings.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+        Input array
+    sep : str or unicode
+        Right-most separator to split each element in array.
+
+    Returns
+    -------
+    out : ndarray
+        Output array of string or unicode, depending on input
+        type.  The output array will have an extra dimension with
+        3 elements per input element.
+
+    See Also
+    --------
+    str.rpartition
+
+    """
+    return _to_bytes_or_str_array(
+        _vec_string(a, object_, 'rpartition', (sep,)), a)
+
+
+def _split_dispatcher(a, sep=None, maxsplit=None):
+    return (a,)
+
+
+@array_function_dispatch(_split_dispatcher)
+def rsplit(a, sep=None, maxsplit=None):
+    """
+    For each element in `a`, return a list of the words in the
+    string, using `sep` as the delimiter string.
+
+    Calls `str.rsplit` element-wise.
+
+    Except for splitting from the right, `rsplit`
+    behaves like `split`.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    sep : str or unicode, optional
+        If `sep` is not specified or None, any whitespace string
+        is a separator.
+    maxsplit : int, optional
+        If `maxsplit` is given, at most `maxsplit` splits are done,
+        the rightmost ones.
+
+    Returns
+    -------
+    out : ndarray
+       Array of list objects
+
+    See Also
+    --------
+    str.rsplit, split
+
+    """
+    # This will return an array of lists of different sizes, so we
+    # leave it as an object array
+    return _vec_string(
+        a, object_, 'rsplit', [sep] + _clean_args(maxsplit))
+
+
+def _strip_dispatcher(a, chars=None):
+    return (a,)
+
+
+@array_function_dispatch(_strip_dispatcher)
+def rstrip(a, chars=None):
+    """
+    For each element in `a`, return a copy with the trailing
+    characters removed.
+
+    Calls `str.rstrip` element-wise.
+
+    Parameters
+    ----------
+    a : array-like of str or unicode
+
+    chars : str or unicode, optional
+       The `chars` argument is a string specifying the set of
+       characters to be removed. If omitted or None, the `chars`
+       argument defaults to removing whitespace. The `chars` argument
+       is not a suffix; rather, all combinations of its values are
+       stripped.
+
+    Returns
+    -------
+    out : ndarray
+        Output array of str or unicode, depending on input type
+
+    See Also
+    --------
+    str.rstrip
+
+    Examples
+    --------
+    >>> c = np.array(['aAaAaA', 'abBABba'], dtype='S7'); c
+    array(['aAaAaA', 'abBABba'],
+        dtype='|S7')
+    >>> np.char.rstrip(c, b'a')
+    array(['aAaAaA', 'abBABb'],
+        dtype='|S7')
+    >>> np.char.rstrip(c, b'A')
+    array(['aAaAa', 'abBABba'],
+        dtype='|S7')
+
+    """
+    a_arr = numpy.asarray(a)
+    return _vec_string(a_arr, a_arr.dtype, 'rstrip', (chars,))
+
+
+@array_function_dispatch(_split_dispatcher)
+def split(a, sep=None, maxsplit=None):
+    """
+    For each element in `a`, return a list of the words in the
+    string, using `sep` as the delimiter string.
+
+    Calls `str.split` element-wise.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    sep : str or unicode, optional
+       If `sep` is not specified or None, any whitespace string is a
+       separator.
+
+    maxsplit : int, optional
+        If `maxsplit` is given, at most `maxsplit` splits are done.
+
+    Returns
+    -------
+    out : ndarray
+        Array of list objects
+
+    See Also
+    --------
+    str.split, rsplit
+
+    """
+    # This will return an array of lists of different sizes, so we
+    # leave it as an object array
+    return _vec_string(
+        a, object_, 'split', [sep] + _clean_args(maxsplit))
+
+
+def _splitlines_dispatcher(a, keepends=None):
+    return (a,)
+
+
+@array_function_dispatch(_splitlines_dispatcher)
+def splitlines(a, keepends=None):
+    """
+    For each element in `a`, return a list of the lines in the
+    element, breaking at line boundaries.
+
+    Calls `str.splitlines` element-wise.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    keepends : bool, optional
+        Line breaks are not included in the resulting list unless
+        keepends is given and true.
+
+    Returns
+    -------
+    out : ndarray
+        Array of list objects
+
+    See Also
+    --------
+    str.splitlines
+
+    """
+    return _vec_string(
+        a, object_, 'splitlines', _clean_args(keepends))
+
+
+def _startswith_dispatcher(a, prefix, start=None, end=None):
+    return (a,)
+
+
+@array_function_dispatch(_startswith_dispatcher)
+def startswith(a, prefix, start=0, end=None):
+    """
+    Returns a boolean array which is `True` where the string element
+    in `a` starts with `prefix`, otherwise `False`.
+
+    Calls `str.startswith` element-wise.
+
+    Parameters
+    ----------
+    a : array_like of str or unicode
+
+    prefix : str
+
+    start, end : int, optional
+        With optional `start`, test beginning at that position. With
+        optional `end`, stop comparing at that position.
+
+    Returns
+    -------
+    out : ndarray
+        Array of booleans
+
+    See Also
+    --------
+    str.startswith
+
+    """
+    return _vec_string(
+        a, bool_, 'startswith', [prefix, start] + _clean_args(end))
+
+
+@array_function_dispatch(_strip_dispatcher)
+def strip(a, chars=None):
+    """
+    For each element in `a`, return a copy with the leading and
+    trailing characters removed.
+
+    Calls `str.strip` element-wise.
+
+    Parameters
+    ----------
+    a : array-like of str or unicode
+
+    chars : str or unicode, optional
+       The `chars` argument is a string specifying the set of
+       characters to be removed. If omitted or None, the `chars`
+       argument defaults to removing whitespace. The `chars` argument
+       is not a prefix or suffix; rather, all combinations of its
+       values are stripped.
+
+    Returns
+    -------
+    out : ndarray
+        Output array of str or unicode, depending on input type
+
+    See Also
+    --------
+    str.strip
+
+    Examples
+    --------
+    >>> c = np.array(['aAaAaA', '  aA  ', 'abBABba'])
+    >>> c
+    array(['aAaAaA', '  aA  ', 'abBABba'], dtype='<U7')
+    >>> np.char.strip(c)
+    array(['aAaAaA', 'aA', 'abBABba'], dtype='<U7')
+    >>> np.char.strip(c, 'a') # 'a' unstripped from c[1] because whitespace leads
+    array(['AaAaA', '  aA  ', 'bBABb'], dtype='<U7')
+    >>> np.char.strip(c, 'A') # 'A' unstripped from c[1] because (unprinted) ws trails
+    array(['aAaAa', '  aA  ', 'abBABba'], dtype='<U7')
+
+    """
+    a_arr = numpy.asarray(a)
+    return _vec_string(a_arr, a_arr.dtype, 'strip', _clean_args(chars))
+
+
+@array_function_dispatch(_unary_op_dispatcher)
+def swapcase(a):
+    """
+    Return element-wise a copy of the string with
+    uppercase characters converted to lowercase and vice versa.
+
+    Calls `str.swapcase` element-wise.
+
+    For 8-bit strings, this method is locale-dependent.
+
+    Parameters
+    ----------
+    a : array_like, {str, unicode}
+        Input array.
+
+    Returns
+    -------
+    out : ndarray, {str, unicode}
+        Output array of str or unicode, depending on input type
+
+    See Also
+    --------
+    str.swapcase
+
+    Examples
+    --------
+    >>> c=np.array(['a1B c','1b Ca','b Ca1','cA1b'],'S5'); c
+    array(['a1B c', '1b Ca', 'b Ca1', 'cA1b'],
+        dtype='|S5')
+    >>> np.char.swapcase(c)
+    array(['A1b C', '1B cA', 'B cA1', 'Ca1B'],
+        dtype='|S5')
+
+    """
+    a_arr = numpy.asarray(a)
+    return _vec_string(a_arr, a_arr.dtype, 'swapcase')
+
+
+@array_function_dispatch(_unary_op_dispatcher)
+def title(a):
+    """
+    Return element-wise title cased version of string or unicode.
+
+    Title case words start with uppercase characters, all remaining cased
+    characters are lowercase.
+
+    Calls `str.title` element-wise.
+
+    For 8-bit strings, this method is locale-dependent.
+
+    Parameters
+    ----------
+    a : array_like, {str, unicode}
+        Input array.
+
+    Returns
+    -------
+    out : ndarray
+        Output array of str or unicode, depending on input type
+
+    See Also
+    --------
+    str.title
+
+    Examples
+    --------
+    >>> c=np.array(['a1b c','1b ca','b ca1','ca1b'],'S5'); c
+    array(['a1b c', '1b ca', 'b ca1', 'ca1b'],
+        dtype='|S5')
+    >>> np.char.title(c)
+    array(['A1B C', '1B Ca', 'B Ca1', 'Ca1B'],
+        dtype='|S5')
+
+    """
+    a_arr = numpy.asarray(a)
+    return _vec_string(a_arr, a_arr.dtype, 'title')
+
+
+def _translate_dispatcher(a, table, deletechars=None):
+    return (a,)
+
+
+@array_function_dispatch(_translate_dispatcher)
+def translate(a, table, deletechars=None):
+    """
+    For each element in `a`, return a copy of the string where all
+    characters occurring in the optional argument `deletechars` are
+    removed, and the remaining characters have been mapped through the
+    given translation table.
+
+    Calls `str.translate` element-wise.
+
+    Parameters
+    ----------
+    a : array-like of str or unicode
+
+    table : str of length 256
+
+    deletechars : str
+
+    Returns
+    -------
+    out : ndarray
+        Output array of str or unicode, depending on input type
+
+    See Also
+    --------
+    str.translate
+
+    """
+    a_arr = numpy.asarray(a)
+    if issubclass(a_arr.dtype.type, str_):
+        return _vec_string(
+            a_arr, a_arr.dtype, 'translate', (table,))
+    else:
+        return _vec_string(
+            a_arr, a_arr.dtype, 'translate', [table] + _clean_args(deletechars))
+
+
+@array_function_dispatch(_unary_op_dispatcher)
+def upper(a):
+    """
+    Return an array with the elements converted to uppercase.
+
+    Calls `str.upper` element-wise.
+
+    For 8-bit strings, this method is locale-dependent.
+
+    Parameters
+    ----------
+    a : array_like, {str, unicode}
+        Input array.
+
+    Returns
+    -------
+    out : ndarray, {str, unicode}
+        Output array of str or unicode, depending on input type
+
+    See Also
+    --------
+    str.upper
+
+    Examples
+    --------
+    >>> c = np.array(['a1b c', '1bca', 'bca1']); c
+    array(['a1b c', '1bca', 'bca1'], dtype='<U5')
+    >>> np.char.upper(c)
+    array(['A1B C', '1BCA', 'BCA1'], dtype='<U5')
+
+    """
+    a_arr = numpy.asarray(a)
+    return _vec_string(a_arr, a_arr.dtype, 'upper')
+
+
+def _zfill_dispatcher(a, width):
+    return (a,)
+
+
+@array_function_dispatch(_zfill_dispatcher)
+def zfill(a, width):
+    """
+    Return the numeric string left-filled with zeros
+
+    Calls `str.zfill` element-wise.
+
+    Parameters
+    ----------
+    a : array_like, {str, unicode}
+        Input array.
+    width : int
+        Width of string to left-fill elements in `a`.
+
+    Returns
+    -------
+    out : ndarray, {str, unicode}
+        Output array of str or unicode, depending on input type
+
+    See Also
+    --------
+    str.zfill
+
+    """
+    a_arr = numpy.asarray(a)
+    width_arr = numpy.asarray(width)
+    size = int(numpy.max(width_arr.flat))
+    return _vec_string(
+        a_arr, type(a_arr.dtype)(size), 'zfill', (width_arr,))
+
+
+@array_function_dispatch(_unary_op_dispatcher)
+def isnumeric(a):
+    """
+    For each element, return True if there are only numeric
+    characters in the element.
+
+    Calls `str.isnumeric` element-wise.
+
+    Numeric characters include digit characters, and all characters
+    that have the Unicode numeric value property, e.g. ``U+2155,
+    VULGAR FRACTION ONE FIFTH``.
+
+    Parameters
+    ----------
+    a : array_like, unicode
+        Input array.
+
+    Returns
+    -------
+    out : ndarray, bool
+        Array of booleans of same shape as `a`.
+
+    See Also
+    --------
+    str.isnumeric
+
+    Examples
+    --------
+    >>> np.char.isnumeric(['123', '123abc', '9.0', '1/4', 'VIII'])
+    array([ True, False, False, False, False])
+
+    """
+    if not _is_unicode(a):
+        raise TypeError("isnumeric is only available for Unicode strings and arrays")
+    return _vec_string(a, bool_, 'isnumeric')
+
+
+@array_function_dispatch(_unary_op_dispatcher)
+def isdecimal(a):
+    """
+    For each element, return True if there are only decimal
+    characters in the element.
+
+    Calls `str.isdecimal` element-wise.
+
+    Decimal characters include digit characters, and all characters
+    that can be used to form decimal-radix numbers,
+    e.g. ``U+0660, ARABIC-INDIC DIGIT ZERO``.
+
+    Parameters
+    ----------
+    a : array_like, unicode
+        Input array.
+
+    Returns
+    -------
+    out : ndarray, bool
+        Array of booleans identical in shape to `a`.
+
+    See Also
+    --------
+    str.isdecimal
+
+    Examples
+    --------
+    >>> np.char.isdecimal(['12345', '4.99', '123ABC', ''])
+    array([ True, False, False, False])
+
+    """ 
+    if not _is_unicode(a):
+        raise TypeError(
+            "isdecimal is only available for Unicode strings and arrays")
+    return _vec_string(a, bool_, 'isdecimal')
+
+
+@set_module('numpy')
+class chararray(ndarray):
+    """
+    chararray(shape, itemsize=1, unicode=False, buffer=None, offset=0,
+              strides=None, order=None)
+
+    Provides a convenient view on arrays of string and unicode values.
+
+    .. note::
+       The `chararray` class exists for backwards compatibility with
+       Numarray, it is not recommended for new development. Starting from numpy
+       1.4, if one needs arrays of strings, it is recommended to use arrays of
+       `dtype` `object_`, `bytes_` or `str_`, and use the free functions
+       in the `numpy.char` module for fast vectorized string operations.
+
+    Versus a regular NumPy array of type `str` or `unicode`, this
+    class adds the following functionality:
+
+      1) values automatically have whitespace removed from the end
+         when indexed
+
+      2) comparison operators automatically remove whitespace from the
+         end when comparing values
+
+      3) vectorized string operations are provided as methods
+         (e.g. `.endswith`) and infix operators (e.g. ``"+", "*", "%"``)
+
+    chararrays should be created using `numpy.char.array` or
+    `numpy.char.asarray`, rather than this constructor directly.
+
+    This constructor creates the array, using `buffer` (with `offset`
+    and `strides`) if it is not ``None``. If `buffer` is ``None``, then
+    constructs a new array with `strides` in "C order", unless both
+    ``len(shape) >= 2`` and ``order='F'``, in which case `strides`
+    is in "Fortran order".
+
+    Methods
+    -------
+    astype
+    argsort
+    copy
+    count
+    decode
+    dump
+    dumps
+    encode
+    endswith
+    expandtabs
+    fill
+    find
+    flatten
+    getfield
+    index
+    isalnum
+    isalpha
+    isdecimal
+    isdigit
+    islower
+    isnumeric
+    isspace
+    istitle
+    isupper
+    item
+    join
+    ljust
+    lower
+    lstrip
+    nonzero
+    put
+    ravel
+    repeat
+    replace
+    reshape
+    resize
+    rfind
+    rindex
+    rjust
+    rsplit
+    rstrip
+    searchsorted
+    setfield
+    setflags
+    sort
+    split
+    splitlines
+    squeeze
+    startswith
+    strip
+    swapaxes
+    swapcase
+    take
+    title
+    tofile
+    tolist
+    tostring
+    translate
+    transpose
+    upper
+    view
+    zfill
+
+    Parameters
+    ----------
+    shape : tuple
+        Shape of the array.
+    itemsize : int, optional
+        Length of each array element, in number of characters. Default is 1.
+    unicode : bool, optional
+        Are the array elements of type unicode (True) or string (False).
+        Default is False.
+    buffer : object exposing the buffer interface or str, optional
+        Memory address of the start of the array data.  Default is None,
+        in which case a new array is created.
+    offset : int, optional
+        Fixed stride displacement from the beginning of an axis?
+        Default is 0. Needs to be >=0.
+    strides : array_like of ints, optional
+        Strides for the array (see `ndarray.strides` for full description).
+        Default is None.
+    order : {'C', 'F'}, optional
+        The order in which the array data is stored in memory: 'C' ->
+        "row major" order (the default), 'F' -> "column major"
+        (Fortran) order.
+
+    Examples
+    --------
+    >>> charar = np.chararray((3, 3))
+    >>> charar[:] = 'a'
+    >>> charar
+    chararray([[b'a', b'a', b'a'],
+               [b'a', b'a', b'a'],
+               [b'a', b'a', b'a']], dtype='|S1')
+
+    >>> charar = np.chararray(charar.shape, itemsize=5)
+    >>> charar[:] = 'abc'
+    >>> charar
+    chararray([[b'abc', b'abc', b'abc'],
+               [b'abc', b'abc', b'abc'],
+               [b'abc', b'abc', b'abc']], dtype='|S5')
+
+    """
+    def __new__(subtype, shape, itemsize=1, unicode=False, buffer=None,
+                offset=0, strides=None, order='C'):
+        global _globalvar
+
+        if unicode:
+            dtype = str_
+        else:
+            dtype = bytes_
+
+        # force itemsize to be a Python int, since using NumPy integer
+        # types results in itemsize.itemsize being used as the size of
+        # strings in the new array.
+        itemsize = int(itemsize)
+
+        if isinstance(buffer, str):
+            # unicode objects do not have the buffer interface
+            filler = buffer
+            buffer = None
+        else:
+            filler = None
+
+        _globalvar = 1
+        if buffer is None:
+            self = ndarray.__new__(subtype, shape, (dtype, itemsize),
+                                   order=order)
+        else:
+            self = ndarray.__new__(subtype, shape, (dtype, itemsize),
+                                   buffer=buffer,
+                                   offset=offset, strides=strides,
+                                   order=order)
+        if filler is not None:
+            self[...] = filler
+        _globalvar = 0
+        return self
+
+    def __array_finalize__(self, obj):
+        # The b is a special case because it is used for reconstructing.
+        if not _globalvar and self.dtype.char not in 'SUbc':
+            raise ValueError("Can only create a chararray from string data.")
+
+    def __getitem__(self, obj):
+        val = ndarray.__getitem__(self, obj)
+
+        if isinstance(val, character):
+            temp = val.rstrip()
+            if len(temp) == 0:
+                val = ''
+            else:
+                val = temp
+
+        return val
+
+    # IMPLEMENTATION NOTE: Most of the methods of this class are
+    # direct delegations to the free functions in this module.
+    # However, those that return an array of strings should instead
+    # return a chararray, so some extra wrapping is required.
+
+    def __eq__(self, other):
+        """
+        Return (self == other) element-wise.
+
+        See Also
+        --------
+        equal
+        """
+        return equal(self, other)
+
+    def __ne__(self, other):
+        """
+        Return (self != other) element-wise.
+
+        See Also
+        --------
+        not_equal
+        """
+        return not_equal(self, other)
+
+    def __ge__(self, other):
+        """
+        Return (self >= other) element-wise.
+
+        See Also
+        --------
+        greater_equal
+        """
+        return greater_equal(self, other)
+
+    def __le__(self, other):
+        """
+        Return (self <= other) element-wise.
+
+        See Also
+        --------
+        less_equal
+        """
+        return less_equal(self, other)
+
+    def __gt__(self, other):
+        """
+        Return (self > other) element-wise.
+
+        See Also
+        --------
+        greater
+        """
+        return greater(self, other)
+
+    def __lt__(self, other):
+        """
+        Return (self < other) element-wise.
+
+        See Also
+        --------
+        less
+        """
+        return less(self, other)
+
+    def __add__(self, other):
+        """
+        Return (self + other), that is string concatenation,
+        element-wise for a pair of array_likes of str or unicode.
+
+        See Also
+        --------
+        add
+        """
+        return asarray(add(self, other))
+
+    def __radd__(self, other):
+        """
+        Return (other + self), that is string concatenation,
+        element-wise for a pair of array_likes of `bytes_` or `str_`.
+
+        See Also
+        --------
+        add
+        """
+        return asarray(add(numpy.asarray(other), self))
+
+    def __mul__(self, i):
+        """
+        Return (self * i), that is string multiple concatenation,
+        element-wise.
+
+        See Also
+        --------
+        multiply
+        """
+        return asarray(multiply(self, i))
+
+    def __rmul__(self, i):
+        """
+        Return (self * i), that is string multiple concatenation,
+        element-wise.
+
+        See Also
+        --------
+        multiply
+        """
+        return asarray(multiply(self, i))
+
+    def __mod__(self, i):
+        """
+        Return (self % i), that is pre-Python 2.6 string formatting
+        (interpolation), element-wise for a pair of array_likes of `bytes_`
+        or `str_`.
+
+        See Also
+        --------
+        mod
+        """
+        return asarray(mod(self, i))
+
+    def __rmod__(self, other):
+        return NotImplemented
+
+    def argsort(self, axis=-1, kind=None, order=None):
+        """
+        Return the indices that sort the array lexicographically.
+
+        For full documentation see `numpy.argsort`, for which this method is
+        in fact merely a "thin wrapper."
+
+        Examples
+        --------
+        >>> c = np.array(['a1b c', '1b ca', 'b ca1', 'Ca1b'], 'S5')
+        >>> c = c.view(np.chararray); c
+        chararray(['a1b c', '1b ca', 'b ca1', 'Ca1b'],
+              dtype='|S5')
+        >>> c[c.argsort()]
+        chararray(['1b ca', 'Ca1b', 'a1b c', 'b ca1'],
+              dtype='|S5')
+
+        """
+        return self.__array__().argsort(axis, kind, order)
+    argsort.__doc__ = ndarray.argsort.__doc__
+
+    def capitalize(self):
+        """
+        Return a copy of `self` with only the first character of each element
+        capitalized.
+
+        See Also
+        --------
+        char.capitalize
+
+        """
+        return asarray(capitalize(self))
+
+    def center(self, width, fillchar=' '):
+        """
+        Return a copy of `self` with its elements centered in a
+        string of length `width`.
+
+        See Also
+        --------
+        center
+        """
+        return asarray(center(self, width, fillchar))
+
+    def count(self, sub, start=0, end=None):
+        """
+        Returns an array with the number of non-overlapping occurrences of
+        substring `sub` in the range [`start`, `end`].
+
+        See Also
+        --------
+        char.count
+
+        """
+        return count(self, sub, start, end)
+
+    def decode(self, encoding=None, errors=None):
+        """
+        Calls ``bytes.decode`` element-wise.
+
+        See Also
+        --------
+        char.decode
+
+        """
+        return decode(self, encoding, errors)
+
+    def encode(self, encoding=None, errors=None):
+        """
+        Calls `str.encode` element-wise.
+
+        See Also
+        --------
+        char.encode
+
+        """
+        return encode(self, encoding, errors)
+
+    def endswith(self, suffix, start=0, end=None):
+        """
+        Returns a boolean array which is `True` where the string element
+        in `self` ends with `suffix`, otherwise `False`.
+
+        See Also
+        --------
+        char.endswith
+
+        """
+        return endswith(self, suffix, start, end)
+
+    def expandtabs(self, tabsize=8):
+        """
+        Return a copy of each string element where all tab characters are
+        replaced by one or more spaces.
+
+        See Also
+        --------
+        char.expandtabs
+
+        """
+        return asarray(expandtabs(self, tabsize))
+
+    def find(self, sub, start=0, end=None):
+        """
+        For each element, return the lowest index in the string where
+        substring `sub` is found.
+
+        See Also
+        --------
+        char.find
+
+        """
+        return find(self, sub, start, end)
+
+    def index(self, sub, start=0, end=None):
+        """
+        Like `find`, but raises `ValueError` when the substring is not found.
+
+        See Also
+        --------
+        char.index
+
+        """
+        return index(self, sub, start, end)
+
+    def isalnum(self):
+        """
+        Returns true for each element if all characters in the string
+        are alphanumeric and there is at least one character, false
+        otherwise.
+
+        See Also
+        --------
+        char.isalnum
+
+        """
+        return isalnum(self)
+
+    def isalpha(self):
+        """
+        Returns true for each element if all characters in the string
+        are alphabetic and there is at least one character, false
+        otherwise.
+
+        See Also
+        --------
+        char.isalpha
+
+        """
+        return isalpha(self)
+
+    def isdigit(self):
+        """
+        Returns true for each element if all characters in the string are
+        digits and there is at least one character, false otherwise.
+
+        See Also
+        --------
+        char.isdigit
+
+        """
+        return isdigit(self)
+
+    def islower(self):
+        """
+        Returns true for each element if all cased characters in the
+        string are lowercase and there is at least one cased character,
+        false otherwise.
+
+        See Also
+        --------
+        char.islower
+
+        """
+        return islower(self)
+
+    def isspace(self):
+        """
+        Returns true for each element if there are only whitespace
+        characters in the string and there is at least one character,
+        false otherwise.
+
+        See Also
+        --------
+        char.isspace
+
+        """
+        return isspace(self)
+
+    def istitle(self):
+        """
+        Returns true for each element if the element is a titlecased
+        string and there is at least one character, false otherwise.
+
+        See Also
+        --------
+        char.istitle
+
+        """
+        return istitle(self)
+
+    def isupper(self):
+        """
+        Returns true for each element if all cased characters in the
+        string are uppercase and there is at least one character, false
+        otherwise.
+
+        See Also
+        --------
+        char.isupper
+
+        """
+        return isupper(self)
+
+    def join(self, seq):
+        """
+        Return a string which is the concatenation of the strings in the
+        sequence `seq`.
+
+        See Also
+        --------
+        char.join
+
+        """
+        return join(self, seq)
+
+    def ljust(self, width, fillchar=' '):
+        """
+        Return an array with the elements of `self` left-justified in a
+        string of length `width`.
+
+        See Also
+        --------
+        char.ljust
+
+        """
+        return asarray(ljust(self, width, fillchar))
+
+    def lower(self):
+        """
+        Return an array with the elements of `self` converted to
+        lowercase.
+
+        See Also
+        --------
+        char.lower
+
+        """
+        return asarray(lower(self))
+
+    def lstrip(self, chars=None):
+        """
+        For each element in `self`, return a copy with the leading characters
+        removed.
+
+        See Also
+        --------
+        char.lstrip
+
+        """
+        return asarray(lstrip(self, chars))
+
+    def partition(self, sep):
+        """
+        Partition each element in `self` around `sep`.
+
+        See Also
+        --------
+        partition
+        """
+        return asarray(partition(self, sep))
+
+    def replace(self, old, new, count=None):
+        """
+        For each element in `self`, return a copy of the string with all
+        occurrences of substring `old` replaced by `new`.
+
+        See Also
+        --------
+        char.replace
+
+        """
+        return asarray(replace(self, old, new, count))
+
+    def rfind(self, sub, start=0, end=None):
+        """
+        For each element in `self`, return the highest index in the string
+        where substring `sub` is found, such that `sub` is contained
+        within [`start`, `end`].
+
+        See Also
+        --------
+        char.rfind
+
+        """
+        return rfind(self, sub, start, end)
+
+    def rindex(self, sub, start=0, end=None):
+        """
+        Like `rfind`, but raises `ValueError` when the substring `sub` is
+        not found.
+
+        See Also
+        --------
+        char.rindex
+
+        """
+        return rindex(self, sub, start, end)
+
+    def rjust(self, width, fillchar=' '):
+        """
+        Return an array with the elements of `self`
+        right-justified in a string of length `width`.
+
+        See Also
+        --------
+        char.rjust
+
+        """
+        return asarray(rjust(self, width, fillchar))
+
+    def rpartition(self, sep):
+        """
+        Partition each element in `self` around `sep`.
+
+        See Also
+        --------
+        rpartition
+        """
+        return asarray(rpartition(self, sep))
+
+    def rsplit(self, sep=None, maxsplit=None):
+        """
+        For each element in `self`, return a list of the words in
+        the string, using `sep` as the delimiter string.
+
+        See Also
+        --------
+        char.rsplit
+
+        """
+        return rsplit(self, sep, maxsplit)
+
+    def rstrip(self, chars=None):
+        """
+        For each element in `self`, return a copy with the trailing
+        characters removed.
+
+        See Also
+        --------
+        char.rstrip
+
+        """
+        return asarray(rstrip(self, chars))
+
+    def split(self, sep=None, maxsplit=None):
+        """
+        For each element in `self`, return a list of the words in the
+        string, using `sep` as the delimiter string.
+
+        See Also
+        --------
+        char.split
+
+        """
+        return split(self, sep, maxsplit)
+
+    def splitlines(self, keepends=None):
+        """
+        For each element in `self`, return a list of the lines in the
+        element, breaking at line boundaries.
+
+        See Also
+        --------
+        char.splitlines
+
+        """
+        return splitlines(self, keepends)
+
+    def startswith(self, prefix, start=0, end=None):
+        """
+        Returns a boolean array which is `True` where the string element
+        in `self` starts with `prefix`, otherwise `False`.
+
+        See Also
+        --------
+        char.startswith
+
+        """
+        return startswith(self, prefix, start, end)
+
+    def strip(self, chars=None):
+        """
+        For each element in `self`, return a copy with the leading and
+        trailing characters removed.
+
+        See Also
+        --------
+        char.strip
+
+        """
+        return asarray(strip(self, chars))
+
+    def swapcase(self):
+        """
+        For each element in `self`, return a copy of the string with
+        uppercase characters converted to lowercase and vice versa.
+
+        See Also
+        --------
+        char.swapcase
+
+        """
+        return asarray(swapcase(self))
+
+    def title(self):
+        """
+        For each element in `self`, return a titlecased version of the
+        string: words start with uppercase characters, all remaining cased
+        characters are lowercase.
+
+        See Also
+        --------
+        char.title
+
+        """
+        return asarray(title(self))
+
+    def translate(self, table, deletechars=None):
+        """
+        For each element in `self`, return a copy of the string where
+        all characters occurring in the optional argument
+        `deletechars` are removed, and the remaining characters have
+        been mapped through the given translation table.
+
+        See Also
+        --------
+        char.translate
+
+        """
+        return asarray(translate(self, table, deletechars))
+
+    def upper(self):
+        """
+        Return an array with the elements of `self` converted to
+        uppercase.
+
+        See Also
+        --------
+        char.upper
+
+        """
+        return asarray(upper(self))
+
+    def zfill(self, width):
+        """
+        Return the numeric string left-filled with zeros in a string of
+        length `width`.
+
+        See Also
+        --------
+        char.zfill
+
+        """
+        return asarray(zfill(self, width))
+
+    def isnumeric(self):
+        """
+        For each element in `self`, return True if there are only
+        numeric characters in the element.
+
+        See Also
+        --------
+        char.isnumeric
+
+        """
+        return isnumeric(self)
+
+    def isdecimal(self):
+        """
+        For each element in `self`, return True if there are only
+        decimal characters in the element.
+
+        See Also
+        --------
+        char.isdecimal
+
+        """
+        return isdecimal(self)
+
+
+@set_module("numpy.char")
+def array(obj, itemsize=None, copy=True, unicode=None, order=None):
+    """
+    Create a `chararray`.
+
+    .. note::
+       This class is provided for numarray backward-compatibility.
+       New code (not concerned with numarray compatibility) should use
+       arrays of type `bytes_` or `str_` and use the free functions
+       in :mod:`numpy.char <numpy.core.defchararray>` for fast
+       vectorized string operations instead.
+
+    Versus a regular NumPy array of type `str` or `unicode`, this
+    class adds the following functionality:
+
+      1) values automatically have whitespace removed from the end
+         when indexed
+
+      2) comparison operators automatically remove whitespace from the
+         end when comparing values
+
+      3) vectorized string operations are provided as methods
+         (e.g. `str.endswith`) and infix operators (e.g. ``+, *, %``)
+
+    Parameters
+    ----------
+    obj : array of str or unicode-like
+
+    itemsize : int, optional
+        `itemsize` is the number of characters per scalar in the
+        resulting array.  If `itemsize` is None, and `obj` is an
+        object array or a Python list, the `itemsize` will be
+        automatically determined.  If `itemsize` is provided and `obj`
+        is of type str or unicode, then the `obj` string will be
+        chunked into `itemsize` pieces.
+
+    copy : bool, optional
+        If true (default), then the object is copied.  Otherwise, a copy
+        will only be made if __array__ returns a copy, if obj is a
+        nested sequence, or if a copy is needed to satisfy any of the other
+        requirements (`itemsize`, unicode, `order`, etc.).
+
+    unicode : bool, optional
+        When true, the resulting `chararray` can contain Unicode
+        characters, when false only 8-bit characters.  If unicode is
+        None and `obj` is one of the following:
+
+          - a `chararray`,
+          - an ndarray of type `str` or `unicode`
+          - a Python str or unicode object,
+
+        then the unicode setting of the output array will be
+        automatically determined.
+
+    order : {'C', 'F', 'A'}, optional
+        Specify the order of the array.  If order is 'C' (default), then the
+        array will be in C-contiguous order (last-index varies the
+        fastest).  If order is 'F', then the returned array
+        will be in Fortran-contiguous order (first-index varies the
+        fastest).  If order is 'A', then the returned array may
+        be in any order (either C-, Fortran-contiguous, or even
+        discontiguous).
+    """
+    if isinstance(obj, (bytes, str)):
+        if unicode is None:
+            if isinstance(obj, str):
+                unicode = True
+            else:
+                unicode = False
+
+        if itemsize is None:
+            itemsize = len(obj)
+        shape = len(obj) // itemsize
+
+        return chararray(shape, itemsize=itemsize, unicode=unicode,
+                         buffer=obj, order=order)
+
+    if isinstance(obj, (list, tuple)):
+        obj = numpy.asarray(obj)
+
+    if isinstance(obj, ndarray) and issubclass(obj.dtype.type, character):
+        # If we just have a vanilla chararray, create a chararray
+        # view around it.
+        if not isinstance(obj, chararray):
+            obj = obj.view(chararray)
+
+        if itemsize is None:
+            itemsize = obj.itemsize
+            # itemsize is in 8-bit chars, so for Unicode, we need
+            # to divide by the size of a single Unicode character,
+            # which for NumPy is always 4
+            if issubclass(obj.dtype.type, str_):
+                itemsize //= 4
+
+        if unicode is None:
+            if issubclass(obj.dtype.type, str_):
+                unicode = True
+            else:
+                unicode = False
+
+        if unicode:
+            dtype = str_
+        else:
+            dtype = bytes_
+
+        if order is not None:
+            obj = numpy.asarray(obj, order=order)
+        if (copy or
+                (itemsize != obj.itemsize) or
+                (not unicode and isinstance(obj, str_)) or
+                (unicode and isinstance(obj, bytes_))):
+            obj = obj.astype((dtype, int(itemsize)))
+        return obj
+
+    if isinstance(obj, ndarray) and issubclass(obj.dtype.type, object):
+        if itemsize is None:
+            # Since no itemsize was specified, convert the input array to
+            # a list so the ndarray constructor will automatically
+            # determine the itemsize for us.
+            obj = obj.tolist()
+            # Fall through to the default case
+
+    if unicode:
+        dtype = str_
+    else:
+        dtype = bytes_
+
+    if itemsize is None:
+        val = narray(obj, dtype=dtype, order=order, subok=True)
+    else:
+        val = narray(obj, dtype=(dtype, itemsize), order=order, subok=True)
+    return val.view(chararray)
+
+
+@set_module("numpy.char")
+def asarray(obj, itemsize=None, unicode=None, order=None):
+    """
+    Convert the input to a `chararray`, copying the data only if
+    necessary.
+
+    Versus a regular NumPy array of type `str` or `unicode`, this
+    class adds the following functionality:
+
+      1) values automatically have whitespace removed from the end
+         when indexed
+
+      2) comparison operators automatically remove whitespace from the
+         end when comparing values
+
+      3) vectorized string operations are provided as methods
+         (e.g. `str.endswith`) and infix operators (e.g. ``+``, ``*``,``%``)
+
+    Parameters
+    ----------
+    obj : array of str or unicode-like
+
+    itemsize : int, optional
+        `itemsize` is the number of characters per scalar in the
+        resulting array.  If `itemsize` is None, and `obj` is an
+        object array or a Python list, the `itemsize` will be
+        automatically determined.  If `itemsize` is provided and `obj`
+        is of type str or unicode, then the `obj` string will be
+        chunked into `itemsize` pieces.
+
+    unicode : bool, optional
+        When true, the resulting `chararray` can contain Unicode
+        characters, when false only 8-bit characters.  If unicode is
+        None and `obj` is one of the following:
+
+          - a `chararray`,
+          - an ndarray of type `str` or 'unicode`
+          - a Python str or unicode object,
+
+        then the unicode setting of the output array will be
+        automatically determined.
+
+    order : {'C', 'F'}, optional
+        Specify the order of the array.  If order is 'C' (default), then the
+        array will be in C-contiguous order (last-index varies the
+        fastest).  If order is 'F', then the returned array
+        will be in Fortran-contiguous order (first-index varies the
+        fastest).
+    """
+    return array(obj, itemsize, copy=False,
+                 unicode=unicode, order=order)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/core/defchararray.pyi

@@ -0,0 +1,421 @@
+from typing import (
+    Literal as L,
+    overload,
+    TypeVar,
+    Any,
+)
+
+from numpy import (
+    chararray as chararray,
+    dtype,
+    str_,
+    bytes_,
+    int_,
+    bool_,
+    object_,
+    _OrderKACF,
+)
+
+from numpy._typing import (
+    NDArray,
+    _ArrayLikeStr_co as U_co,
+    _ArrayLikeBytes_co as S_co,
+    _ArrayLikeInt_co as i_co,
+    _ArrayLikeBool_co as b_co,
+)
+
+from numpy.core.multiarray import compare_chararrays as compare_chararrays
+
+_SCT = TypeVar("_SCT", str_, bytes_)
+_CharArray = chararray[Any, dtype[_SCT]]
+
+__all__: list[str]
+
+# Comparison
+@overload
+def equal(x1: U_co, x2: U_co) -> NDArray[bool_]: ...
+@overload
+def equal(x1: S_co, x2: S_co) -> NDArray[bool_]: ...
+
+@overload
+def not_equal(x1: U_co, x2: U_co) -> NDArray[bool_]: ...
+@overload
+def not_equal(x1: S_co, x2: S_co) -> NDArray[bool_]: ...
+
+@overload
+def greater_equal(x1: U_co, x2: U_co) -> NDArray[bool_]: ...
+@overload
+def greater_equal(x1: S_co, x2: S_co) -> NDArray[bool_]: ...
+
+@overload
+def less_equal(x1: U_co, x2: U_co) -> NDArray[bool_]: ...
+@overload
+def less_equal(x1: S_co, x2: S_co) -> NDArray[bool_]: ...
+
+@overload
+def greater(x1: U_co, x2: U_co) -> NDArray[bool_]: ...
+@overload
+def greater(x1: S_co, x2: S_co) -> NDArray[bool_]: ...
+
+@overload
+def less(x1: U_co, x2: U_co) -> NDArray[bool_]: ...
+@overload
+def less(x1: S_co, x2: S_co) -> NDArray[bool_]: ...
+
+# String operations
+@overload
+def add(x1: U_co, x2: U_co) -> NDArray[str_]: ...
+@overload
+def add(x1: S_co, x2: S_co) -> NDArray[bytes_]: ...
+
+@overload
+def multiply(a: U_co, i: i_co) -> NDArray[str_]: ...
+@overload
+def multiply(a: S_co, i: i_co) -> NDArray[bytes_]: ...
+
+@overload
+def mod(a: U_co, value: Any) -> NDArray[str_]: ...
+@overload
+def mod(a: S_co, value: Any) -> NDArray[bytes_]: ...
+
+@overload
+def capitalize(a: U_co) -> NDArray[str_]: ...
+@overload
+def capitalize(a: S_co) -> NDArray[bytes_]: ...
+
+@overload
+def center(a: U_co, width: i_co, fillchar: U_co = ...) -> NDArray[str_]: ...
+@overload
+def center(a: S_co, width: i_co, fillchar: S_co = ...) -> NDArray[bytes_]: ...
+
+def decode(
+    a: S_co,
+    encoding: None | str = ...,
+    errors: None | str = ...,
+) -> NDArray[str_]: ...
+
+def encode(
+    a: U_co,
+    encoding: None | str = ...,
+    errors: None | str = ...,
+) -> NDArray[bytes_]: ...
+
+@overload
+def expandtabs(a: U_co, tabsize: i_co = ...) -> NDArray[str_]: ...
+@overload
+def expandtabs(a: S_co, tabsize: i_co = ...) -> NDArray[bytes_]: ...
+
+@overload
+def join(sep: U_co, seq: U_co) -> NDArray[str_]: ...
+@overload
+def join(sep: S_co, seq: S_co) -> NDArray[bytes_]: ...
+
+@overload
+def ljust(a: U_co, width: i_co, fillchar: U_co = ...) -> NDArray[str_]: ...
+@overload
+def ljust(a: S_co, width: i_co, fillchar: S_co = ...) -> NDArray[bytes_]: ...
+
+@overload
+def lower(a: U_co) -> NDArray[str_]: ...
+@overload
+def lower(a: S_co) -> NDArray[bytes_]: ...
+
+@overload
+def lstrip(a: U_co, chars: None | U_co = ...) -> NDArray[str_]: ...
+@overload
+def lstrip(a: S_co, chars: None | S_co = ...) -> NDArray[bytes_]: ...
+
+@overload
+def partition(a: U_co, sep: U_co) -> NDArray[str_]: ...
+@overload
+def partition(a: S_co, sep: S_co) -> NDArray[bytes_]: ...
+
+@overload
+def replace(
+    a: U_co,
+    old: U_co,
+    new: U_co,
+    count: None | i_co = ...,
+) -> NDArray[str_]: ...
+@overload
+def replace(
+    a: S_co,
+    old: S_co,
+    new: S_co,
+    count: None | i_co = ...,
+) -> NDArray[bytes_]: ...
+
+@overload
+def rjust(
+    a: U_co,
+    width: i_co,
+    fillchar: U_co = ...,
+) -> NDArray[str_]: ...
+@overload
+def rjust(
+    a: S_co,
+    width: i_co,
+    fillchar: S_co = ...,
+) -> NDArray[bytes_]: ...
+
+@overload
+def rpartition(a: U_co, sep: U_co) -> NDArray[str_]: ...
+@overload
+def rpartition(a: S_co, sep: S_co) -> NDArray[bytes_]: ...
+
+@overload
+def rsplit(
+    a: U_co,
+    sep: None | U_co = ...,
+    maxsplit: None | i_co = ...,
+) -> NDArray[object_]: ...
+@overload
+def rsplit(
+    a: S_co,
+    sep: None | S_co = ...,
+    maxsplit: None | i_co = ...,
+) -> NDArray[object_]: ...
+
+@overload
+def rstrip(a: U_co, chars: None | U_co = ...) -> NDArray[str_]: ...
+@overload
+def rstrip(a: S_co, chars: None | S_co = ...) -> NDArray[bytes_]: ...
+
+@overload
+def split(
+    a: U_co,
+    sep: None | U_co = ...,
+    maxsplit: None | i_co = ...,
+) -> NDArray[object_]: ...
+@overload
+def split(
+    a: S_co,
+    sep: None | S_co = ...,
+    maxsplit: None | i_co = ...,
+) -> NDArray[object_]: ...
+
+@overload
+def splitlines(a: U_co, keepends: None | b_co = ...) -> NDArray[object_]: ...
+@overload
+def splitlines(a: S_co, keepends: None | b_co = ...) -> NDArray[object_]: ...
+
+@overload
+def strip(a: U_co, chars: None | U_co = ...) -> NDArray[str_]: ...
+@overload
+def strip(a: S_co, chars: None | S_co = ...) -> NDArray[bytes_]: ...
+
+@overload
+def swapcase(a: U_co) -> NDArray[str_]: ...
+@overload
+def swapcase(a: S_co) -> NDArray[bytes_]: ...
+
+@overload
+def title(a: U_co) -> NDArray[str_]: ...
+@overload
+def title(a: S_co) -> NDArray[bytes_]: ...
+
+@overload
+def translate(
+    a: U_co,
+    table: U_co,
+    deletechars: None | U_co = ...,
+) -> NDArray[str_]: ...
+@overload
+def translate(
+    a: S_co,
+    table: S_co,
+    deletechars: None | S_co = ...,
+) -> NDArray[bytes_]: ...
+
+@overload
+def upper(a: U_co) -> NDArray[str_]: ...
+@overload
+def upper(a: S_co) -> NDArray[bytes_]: ...
+
+@overload
+def zfill(a: U_co, width: i_co) -> NDArray[str_]: ...
+@overload
+def zfill(a: S_co, width: i_co) -> NDArray[bytes_]: ...
+
+# String information
+@overload
+def count(
+    a: U_co,
+    sub: U_co,
+    start: i_co = ...,
+    end: None | i_co = ...,
+) -> NDArray[int_]: ...
+@overload
+def count(
+    a: S_co,
+    sub: S_co,
+    start: i_co = ...,
+    end: None | i_co = ...,
+) -> NDArray[int_]: ...
+
+@overload
+def endswith(
+    a: U_co,
+    suffix: U_co,
+    start: i_co = ...,
+    end: None | i_co = ...,
+) -> NDArray[bool_]: ...
+@overload
+def endswith(
+    a: S_co,
+    suffix: S_co,
+    start: i_co = ...,
+    end: None | i_co = ...,
+) -> NDArray[bool_]: ...
+
+@overload
+def find(
+    a: U_co,
+    sub: U_co,
+    start: i_co = ...,
+    end: None | i_co = ...,
+) -> NDArray[int_]: ...
+@overload
+def find(
+    a: S_co,
+    sub: S_co,
+    start: i_co = ...,
+    end: None | i_co = ...,
+) -> NDArray[int_]: ...
+
+@overload
+def index(
+    a: U_co,
+    sub: U_co,
+    start: i_co = ...,
+    end: None | i_co = ...,
+) -> NDArray[int_]: ...
+@overload
+def index(
+    a: S_co,
+    sub: S_co,
+    start: i_co = ...,
+    end: None | i_co = ...,
+) -> NDArray[int_]: ...
+
+def isalpha(a: U_co | S_co) -> NDArray[bool_]: ...
+def isalnum(a: U_co | S_co) -> NDArray[bool_]: ...
+def isdecimal(a: U_co | S_co) -> NDArray[bool_]: ...
+def isdigit(a: U_co | S_co) -> NDArray[bool_]: ...
+def islower(a: U_co | S_co) -> NDArray[bool_]: ...
+def isnumeric(a: U_co | S_co) -> NDArray[bool_]: ...
+def isspace(a: U_co | S_co) -> NDArray[bool_]: ...
+def istitle(a: U_co | S_co) -> NDArray[bool_]: ...
+def isupper(a: U_co | S_co) -> NDArray[bool_]: ...
+
+@overload
+def rfind(
+    a: U_co,
+    sub: U_co,
+    start: i_co = ...,
+    end: None | i_co = ...,
+) -> NDArray[int_]: ...
+@overload
+def rfind(
+    a: S_co,
+    sub: S_co,
+    start: i_co = ...,
+    end: None | i_co = ...,
+) -> NDArray[int_]: ...
+
+@overload
+def rindex(
+    a: U_co,
+    sub: U_co,
+    start: i_co = ...,
+    end: None | i_co = ...,
+) -> NDArray[int_]: ...
+@overload
+def rindex(
+    a: S_co,
+    sub: S_co,
+    start: i_co = ...,
+    end: None | i_co = ...,
+) -> NDArray[int_]: ...
+
+@overload
+def startswith(
+    a: U_co,
+    prefix: U_co,
+    start: i_co = ...,
+    end: None | i_co = ...,
+) -> NDArray[bool_]: ...
+@overload
+def startswith(
+    a: S_co,
+    prefix: S_co,
+    start: i_co = ...,
+    end: None | i_co = ...,
+) -> NDArray[bool_]: ...
+
+def str_len(A: U_co | S_co) -> NDArray[int_]: ...
+
+# Overload 1 and 2: str- or bytes-based array-likes
+# overload 3: arbitrary object with unicode=False  (-> bytes_)
+# overload 4: arbitrary object with unicode=True  (-> str_)
+@overload
+def array(
+    obj: U_co,
+    itemsize: None | int = ...,
+    copy: bool = ...,
+    unicode: L[False] = ...,
+    order: _OrderKACF = ...,
+) -> _CharArray[str_]: ...
+@overload
+def array(
+    obj: S_co,
+    itemsize: None | int = ...,
+    copy: bool = ...,
+    unicode: L[False] = ...,
+    order: _OrderKACF = ...,
+) -> _CharArray[bytes_]: ...
+@overload
+def array(
+    obj: object,
+    itemsize: None | int = ...,
+    copy: bool = ...,
+    unicode: L[False] = ...,
+    order: _OrderKACF = ...,
+) -> _CharArray[bytes_]: ...
+@overload
+def array(
+    obj: object,
+    itemsize: None | int = ...,
+    copy: bool = ...,
+    unicode: L[True] = ...,
+    order: _OrderKACF = ...,
+) -> _CharArray[str_]: ...
+
+@overload
+def asarray(
+    obj: U_co,
+    itemsize: None | int = ...,
+    unicode: L[False] = ...,
+    order: _OrderKACF = ...,
+) -> _CharArray[str_]: ...
+@overload
+def asarray(
+    obj: S_co,
+    itemsize: None | int = ...,
+    unicode: L[False] = ...,
+    order: _OrderKACF = ...,
+) -> _CharArray[bytes_]: ...
+@overload
+def asarray(
+    obj: object,
+    itemsize: None | int = ...,
+    unicode: L[False] = ...,
+    order: _OrderKACF = ...,
+) -> _CharArray[bytes_]: ...
+@overload
+def asarray(
+    obj: object,
+    itemsize: None | int = ...,
+    unicode: L[True] = ...,
+    order: _OrderKACF = ...,
+) -> _CharArray[str_]: ...

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/core/tests/test_argparse.py

@@ -0,0 +1,62 @@
+"""
+Tests for the private NumPy argument parsing functionality.
+They mainly exists to ensure good test coverage without having to try the
+weirder cases on actual numpy functions but test them in one place.
+
+The test function is defined in C to be equivalent to (errors may not always
+match exactly, and could be adjusted):
+
+    def func(arg1, /, arg2, *, arg3):
+        i = integer(arg1)  # reproducing the 'i' parsing in Python.
+        return None
+"""
+
+import pytest
+
+import numpy as np
+from numpy.core._multiarray_tests import argparse_example_function as func
+
+
+def test_invalid_integers():
+    with pytest.raises(TypeError,
+            match="integer argument expected, got float"):
+        func(1.)
+    with pytest.raises(OverflowError):
+        func(2**100)
+
+
+def test_missing_arguments():
+    with pytest.raises(TypeError,
+            match="missing required positional argument 0"):
+        func()
+    with pytest.raises(TypeError,
+            match="missing required positional argument 0"):
+        func(arg2=1, arg3=4)
+    with pytest.raises(TypeError,
+            match=r"missing required argument \'arg2\' \(pos 1\)"):
+        func(1, arg3=5)
+
+
+def test_too_many_positional():
+    # the second argument is positional but can be passed as keyword.
+    with pytest.raises(TypeError,
+            match="takes from 2 to 3 positional arguments but 4 were given"):
+        func(1, 2, 3, 4)
+
+
+def test_multiple_values():
+    with pytest.raises(TypeError,
+            match=r"given by name \('arg2'\) and position \(position 1\)"):
+        func(1, 2, arg2=3)
+
+
+def test_string_fallbacks():
+    # We can (currently?) use numpy strings to test the "slow" fallbacks
+    # that should normally not be taken due to string interning.
+    arg2 = np.str_("arg2")
+    missing_arg = np.str_("missing_arg")
+    func(1, **{arg2: 3})
+    with pytest.raises(TypeError,
+            match="got an unexpected keyword argument 'missing_arg'"):
+        func(2, **{missing_arg: 3})
+

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/core/tests/test_array_coercion.py

@@ -0,0 +1,898 @@
+"""
+Tests for array coercion, mainly through testing `np.array` results directly.
+Note that other such tests exist, e.g., in `test_api.py` and many corner-cases
+are tested (sometimes indirectly) elsewhere.
+"""
+
+from itertools import permutations, product
+
+import pytest
+from pytest import param
+
+import numpy as np
+from numpy.core._rational_tests import rational
+from numpy.core._multiarray_umath import _discover_array_parameters
+
+from numpy.testing import (
+    assert_array_equal, assert_warns, IS_PYPY)
+
+
+def arraylikes():
+    """
+    Generator for functions converting an array into various array-likes.
+    If full is True (default) it includes array-likes not capable of handling
+    all dtypes.
+    """
+    # base array:
+    def ndarray(a):
+        return a
+
+    yield param(ndarray, id="ndarray")
+
+    # subclass:
+    class MyArr(np.ndarray):
+        pass
+
+    def subclass(a):
+        return a.view(MyArr)
+
+    yield subclass
+
+    class _SequenceLike():
+        # Older NumPy versions, sometimes cared whether a protocol array was
+        # also _SequenceLike.  This shouldn't matter, but keep it for now
+        # for __array__ and not the others.
+        def __len__(self):
+            raise TypeError
+
+        def __getitem__(self):
+            raise TypeError
+
+    # Array-interface
+    class ArrayDunder(_SequenceLike):
+        def __init__(self, a):
+            self.a = a
+
+        def __array__(self, dtype=None):
+            return self.a
+
+    yield param(ArrayDunder, id="__array__")
+
+    # memory-view
+    yield param(memoryview, id="memoryview")
+
+    # Array-interface
+    class ArrayInterface:
+        def __init__(self, a):
+            self.a = a  # need to hold on to keep interface valid
+            self.__array_interface__ = a.__array_interface__
+
+    yield param(ArrayInterface, id="__array_interface__")
+
+    # Array-Struct
+    class ArrayStruct:
+        def __init__(self, a):
+            self.a = a  # need to hold on to keep struct valid
+            self.__array_struct__ = a.__array_struct__
+
+    yield param(ArrayStruct, id="__array_struct__")
+
+
+def scalar_instances(times=True, extended_precision=True, user_dtype=True):
+    # Hard-coded list of scalar instances.
+    # Floats:
+    yield param(np.sqrt(np.float16(5)), id="float16")
+    yield param(np.sqrt(np.float32(5)), id="float32")
+    yield param(np.sqrt(np.float64(5)), id="float64")
+    if extended_precision:
+        yield param(np.sqrt(np.longdouble(5)), id="longdouble")
+
+    # Complex:
+    yield param(np.sqrt(np.complex64(2+3j)), id="complex64")
+    yield param(np.sqrt(np.complex128(2+3j)), id="complex128")
+    if extended_precision:
+        yield param(np.sqrt(np.longcomplex(2+3j)), id="clongdouble")
+
+    # Bool:
+    # XFAIL: Bool should be added, but has some bad properties when it
+    # comes to strings, see also gh-9875
+    # yield param(np.bool_(0), id="bool")
+
+    # Integers:
+    yield param(np.int8(2), id="int8")
+    yield param(np.int16(2), id="int16")
+    yield param(np.int32(2), id="int32")
+    yield param(np.int64(2), id="int64")
+
+    yield param(np.uint8(2), id="uint8")
+    yield param(np.uint16(2), id="uint16")
+    yield param(np.uint32(2), id="uint32")
+    yield param(np.uint64(2), id="uint64")
+
+    # Rational:
+    if user_dtype:
+        yield param(rational(1, 2), id="rational")
+
+    # Cannot create a structured void scalar directly:
+    structured = np.array([(1, 3)], "i,i")[0]
+    assert isinstance(structured, np.void)
+    assert structured.dtype == np.dtype("i,i")
+    yield param(structured, id="structured")
+
+    if times:
+        # Datetimes and timedelta
+        yield param(np.timedelta64(2), id="timedelta64[generic]")
+        yield param(np.timedelta64(23, "s"), id="timedelta64[s]")
+        yield param(np.timedelta64("NaT", "s"), id="timedelta64[s](NaT)")
+
+        yield param(np.datetime64("NaT"), id="datetime64[generic](NaT)")
+        yield param(np.datetime64("2020-06-07 12:43", "ms"), id="datetime64[ms]")
+
+    # Strings and unstructured void:
+    yield param(np.bytes_(b"1234"), id="bytes")
+    yield param(np.str_("2345"), id="unicode")
+    yield param(np.void(b"4321"), id="unstructured_void")
+
+
+def is_parametric_dtype(dtype):
+    """Returns True if the dtype is a parametric legacy dtype (itemsize
+    is 0, or a datetime without units)
+    """
+    if dtype.itemsize == 0:
+        return True
+    if issubclass(dtype.type, (np.datetime64, np.timedelta64)):
+        if dtype.name.endswith("64"):
+            # Generic time units
+            return True
+    return False
+
+
+class TestStringDiscovery:
+    @pytest.mark.parametrize("obj",
+            [object(), 1.2, 10**43, None, "string"],
+            ids=["object", "1.2", "10**43", "None", "string"])
+    def test_basic_stringlength(self, obj):
+        length = len(str(obj))
+        expected = np.dtype(f"S{length}")
+
+        assert np.array(obj, dtype="S").dtype == expected
+        assert np.array([obj], dtype="S").dtype == expected
+
+        # A nested array is also discovered correctly
+        arr = np.array(obj, dtype="O")
+        assert np.array(arr, dtype="S").dtype == expected
+        # Also if we use the dtype class
+        assert np.array(arr, dtype=type(expected)).dtype == expected
+        # Check that .astype() behaves identical
+        assert arr.astype("S").dtype == expected
+        # The DType class is accepted by `.astype()`
+        assert arr.astype(type(np.dtype("S"))).dtype == expected
+
+    @pytest.mark.parametrize("obj",
+            [object(), 1.2, 10**43, None, "string"],
+            ids=["object", "1.2", "10**43", "None", "string"])
+    def test_nested_arrays_stringlength(self, obj):
+        length = len(str(obj))
+        expected = np.dtype(f"S{length}")
+        arr = np.array(obj, dtype="O")
+        assert np.array([arr, arr], dtype="S").dtype == expected
+
+    @pytest.mark.parametrize("arraylike", arraylikes())
+    def test_unpack_first_level(self, arraylike):
+        # We unpack exactly one level of array likes
+        obj = np.array([None])
+        obj[0] = np.array(1.2)
+        # the length of the included item, not of the float dtype
+        length = len(str(obj[0]))
+        expected = np.dtype(f"S{length}")
+
+        obj = arraylike(obj)
+        # casting to string usually calls str(obj)
+        arr = np.array([obj], dtype="S")
+        assert arr.shape == (1, 1)
+        assert arr.dtype == expected
+
+
+class TestScalarDiscovery:
+    def test_void_special_case(self):
+        # Void dtypes with structures discover tuples as elements
+        arr = np.array((1, 2, 3), dtype="i,i,i")
+        assert arr.shape == ()
+        arr = np.array([(1, 2, 3)], dtype="i,i,i")
+        assert arr.shape == (1,)
+
+    def test_char_special_case(self):
+        arr = np.array("string", dtype="c")
+        assert arr.shape == (6,)
+        assert arr.dtype.char == "c"
+        arr = np.array(["string"], dtype="c")
+        assert arr.shape == (1, 6)
+        assert arr.dtype.char == "c"
+
+    def test_char_special_case_deep(self):
+        # Check that the character special case errors correctly if the
+        # array is too deep:
+        nested = ["string"]  # 2 dimensions (due to string being sequence)
+        for i in range(np.MAXDIMS - 2):
+            nested = [nested]
+
+        arr = np.array(nested, dtype='c')
+        assert arr.shape == (1,) * (np.MAXDIMS - 1) + (6,)
+        with pytest.raises(ValueError):
+            np.array([nested], dtype="c")
+
+    def test_unknown_object(self):
+        arr = np.array(object())
+        assert arr.shape == ()
+        assert arr.dtype == np.dtype("O")
+
+    @pytest.mark.parametrize("scalar", scalar_instances())
+    def test_scalar(self, scalar):
+        arr = np.array(scalar)
+        assert arr.shape == ()
+        assert arr.dtype == scalar.dtype
+
+        arr = np.array([[scalar, scalar]])
+        assert arr.shape == (1, 2)
+        assert arr.dtype == scalar.dtype
+
+    # Additionally to string this test also runs into a corner case
+    # with datetime promotion (the difference is the promotion order).
+    @pytest.mark.filterwarnings("ignore:Promotion of numbers:FutureWarning")
+    def test_scalar_promotion(self):
+        for sc1, sc2 in product(scalar_instances(), scalar_instances()):
+            sc1, sc2 = sc1.values[0], sc2.values[0]
+            # test all combinations:
+            try:
+                arr = np.array([sc1, sc2])
+            except (TypeError, ValueError):
+                # The promotion between two times can fail
+                # XFAIL (ValueError): Some object casts are currently undefined
+                continue
+            assert arr.shape == (2,)
+            try:
+                dt1, dt2 = sc1.dtype, sc2.dtype
+                expected_dtype = np.promote_types(dt1, dt2)
+                assert arr.dtype == expected_dtype
+            except TypeError as e:
+                # Will currently always go to object dtype
+                assert arr.dtype == np.dtype("O")
+
+    @pytest.mark.parametrize("scalar", scalar_instances())
+    def test_scalar_coercion(self, scalar):
+        # This tests various scalar coercion paths, mainly for the numerical
+        # types. It includes some paths not directly related to `np.array`.
+        if isinstance(scalar, np.inexact):
+            # Ensure we have a full-precision number if available
+            scalar = type(scalar)((scalar * 2)**0.5)
+
+        if type(scalar) is rational:
+            # Rational generally fails due to a missing cast. In the future
+            # object casts should automatically be defined based on `setitem`.
+            pytest.xfail("Rational to object cast is undefined currently.")
+
+        # Use casting from object:
+        arr = np.array(scalar, dtype=object).astype(scalar.dtype)
+
+        # Test various ways to create an array containing this scalar:
+        arr1 = np.array(scalar).reshape(1)
+        arr2 = np.array([scalar])
+        arr3 = np.empty(1, dtype=scalar.dtype)
+        arr3[0] = scalar
+        arr4 = np.empty(1, dtype=scalar.dtype)
+        arr4[:] = [scalar]
+        # All of these methods should yield the same results
+        assert_array_equal(arr, arr1)
+        assert_array_equal(arr, arr2)
+        assert_array_equal(arr, arr3)
+        assert_array_equal(arr, arr4)
+
+    @pytest.mark.xfail(IS_PYPY, reason="`int(np.complex128(3))` fails on PyPy")
+    @pytest.mark.filterwarnings("ignore::numpy.ComplexWarning")
+    @pytest.mark.parametrize("cast_to", scalar_instances())
+    def test_scalar_coercion_same_as_cast_and_assignment(self, cast_to):
+        """
+        Test that in most cases:
+           * `np.array(scalar, dtype=dtype)`
+           * `np.empty((), dtype=dtype)[()] = scalar`
+           * `np.array(scalar).astype(dtype)`
+        should behave the same.  The only exceptions are parametric dtypes
+        (mainly datetime/timedelta without unit) and void without fields.
+        """
+        dtype = cast_to.dtype  # use to parametrize only the target dtype
+
+        for scalar in scalar_instances(times=False):
+            scalar = scalar.values[0]
+
+            if dtype.type == np.void:
+               if scalar.dtype.fields is not None and dtype.fields is None:
+                    # Here, coercion to "V6" works, but the cast fails.
+                    # Since the types are identical, SETITEM takes care of
+                    # this, but has different rules than the cast.
+                    with pytest.raises(TypeError):
+                        np.array(scalar).astype(dtype)
+                    np.array(scalar, dtype=dtype)
+                    np.array([scalar], dtype=dtype)
+                    continue
+
+            # The main test, we first try to use casting and if it succeeds
+            # continue below testing that things are the same, otherwise
+            # test that the alternative paths at least also fail.
+            try:
+                cast = np.array(scalar).astype(dtype)
+            except (TypeError, ValueError, RuntimeError):
+                # coercion should also raise (error type may change)
+                with pytest.raises(Exception):
+                    np.array(scalar, dtype=dtype)
+
+                if (isinstance(scalar, rational) and
+                        np.issubdtype(dtype, np.signedinteger)):
+                    return
+
+                with pytest.raises(Exception):
+                    np.array([scalar], dtype=dtype)
+                # assignment should also raise
+                res = np.zeros((), dtype=dtype)
+                with pytest.raises(Exception):
+                    res[()] = scalar
+
+                return
+
+            # Non error path:
+            arr = np.array(scalar, dtype=dtype)
+            assert_array_equal(arr, cast)
+            # assignment behaves the same
+            ass = np.zeros((), dtype=dtype)
+            ass[()] = scalar
+            assert_array_equal(ass, cast)
+
+    @pytest.mark.parametrize("pyscalar", [10, 10.32, 10.14j, 10**100])
+    def test_pyscalar_subclasses(self, pyscalar):
+        """NumPy arrays are read/write which means that anything but invariant
+        behaviour is on thin ice.  However, we currently are happy to discover
+        subclasses of Python float, int, complex the same as the base classes.
+        This should potentially be deprecated.
+        """
+        class MyScalar(type(pyscalar)):
+            pass
+
+        res = np.array(MyScalar(pyscalar))
+        expected = np.array(pyscalar)
+        assert_array_equal(res, expected)
+
+    @pytest.mark.parametrize("dtype_char", np.typecodes["All"])
+    def test_default_dtype_instance(self, dtype_char):
+        if dtype_char in "SU":
+            dtype = np.dtype(dtype_char + "1")
+        elif dtype_char == "V":
+            # Legacy behaviour was to use V8. The reason was float64 being the
+            # default dtype and that having 8 bytes.
+            dtype = np.dtype("V8")
+        else:
+            dtype = np.dtype(dtype_char)
+
+        discovered_dtype, _ = _discover_array_parameters([], type(dtype))
+
+        assert discovered_dtype == dtype
+        assert discovered_dtype.itemsize == dtype.itemsize
+
+    @pytest.mark.parametrize("dtype", np.typecodes["Integer"])
+    @pytest.mark.parametrize(["scalar", "error"],
+            [(np.float64(np.nan), ValueError),
+             (np.array(-1).astype(np.ulonglong)[()], OverflowError)])
+    def test_scalar_to_int_coerce_does_not_cast(self, dtype, scalar, error):
+        """
+        Signed integers are currently different in that they do not cast other
+        NumPy scalar, but instead use scalar.__int__(). The hardcoded
+        exception to this rule is `np.array(scalar, dtype=integer)`.
+        """
+        dtype = np.dtype(dtype)
+
+        # This is a special case using casting logic. It warns for the NaN
+        # but allows the cast (giving undefined behaviour).
+        with np.errstate(invalid="ignore"):
+            coerced = np.array(scalar, dtype=dtype)
+            cast = np.array(scalar).astype(dtype)
+        assert_array_equal(coerced, cast)
+
+        # However these fail:
+        with pytest.raises(error):
+            np.array([scalar], dtype=dtype)
+        with pytest.raises(error):
+            cast[()] = scalar
+
+
+class TestTimeScalars:
+    @pytest.mark.parametrize("dtype", [np.int64, np.float32])
+    @pytest.mark.parametrize("scalar",
+            [param(np.timedelta64("NaT", "s"), id="timedelta64[s](NaT)"),
+             param(np.timedelta64(123, "s"), id="timedelta64[s]"),
+             param(np.datetime64("NaT", "generic"), id="datetime64[generic](NaT)"),
+             param(np.datetime64(1, "D"), id="datetime64[D]")],)
+    def test_coercion_basic(self, dtype, scalar):
+        # Note the `[scalar]` is there because np.array(scalar) uses stricter
+        # `scalar.__int__()` rules for backward compatibility right now.
+        arr = np.array(scalar, dtype=dtype)
+        cast = np.array(scalar).astype(dtype)
+        assert_array_equal(arr, cast)
+
+        ass = np.ones((), dtype=dtype)
+        if issubclass(dtype, np.integer):
+            with pytest.raises(TypeError):
+                # raises, as would np.array([scalar], dtype=dtype), this is
+                # conversion from times, but behaviour of integers.
+                ass[()] = scalar
+        else:
+            ass[()] = scalar
+            assert_array_equal(ass, cast)
+
+    @pytest.mark.parametrize("dtype", [np.int64, np.float32])
+    @pytest.mark.parametrize("scalar",
+            [param(np.timedelta64(123, "ns"), id="timedelta64[ns]"),
+             param(np.timedelta64(12, "generic"), id="timedelta64[generic]")])
+    def test_coercion_timedelta_convert_to_number(self, dtype, scalar):
+        # Only "ns" and "generic" timedeltas can be converted to numbers
+        # so these are slightly special.
+        arr = np.array(scalar, dtype=dtype)
+        cast = np.array(scalar).astype(dtype)
+        ass = np.ones((), dtype=dtype)
+        ass[()] = scalar  # raises, as would np.array([scalar], dtype=dtype)
+
+        assert_array_equal(arr, cast)
+        assert_array_equal(cast, cast)
+
+    @pytest.mark.parametrize("dtype", ["S6", "U6"])
+    @pytest.mark.parametrize(["val", "unit"],
+            [param(123, "s", id="[s]"), param(123, "D", id="[D]")])
+    def test_coercion_assignment_datetime(self, val, unit, dtype):
+        # String from datetime64 assignment is currently special cased to
+        # never use casting.  This is because casting will error in this
+        # case, and traditionally in most cases the behaviour is maintained
+        # like this.  (`np.array(scalar, dtype="U6")` would have failed before)
+        # TODO: This discrepancy _should_ be resolved, either by relaxing the
+        #       cast, or by deprecating the first part.
+        scalar = np.datetime64(val, unit)
+        dtype = np.dtype(dtype)
+        cut_string = dtype.type(str(scalar)[:6])
+
+        arr = np.array(scalar, dtype=dtype)
+        assert arr[()] == cut_string
+        ass = np.ones((), dtype=dtype)
+        ass[()] = scalar
+        assert ass[()] == cut_string
+
+        with pytest.raises(RuntimeError):
+            # However, unlike the above assignment using `str(scalar)[:6]`
+            # due to being handled by the string DType and not be casting
+            # the explicit cast fails:
+            np.array(scalar).astype(dtype)
+
+
+    @pytest.mark.parametrize(["val", "unit"],
+            [param(123, "s", id="[s]"), param(123, "D", id="[D]")])
+    def test_coercion_assignment_timedelta(self, val, unit):
+        scalar = np.timedelta64(val, unit)
+
+        # Unlike datetime64, timedelta allows the unsafe cast:
+        np.array(scalar, dtype="S6")
+        cast = np.array(scalar).astype("S6")
+        ass = np.ones((), dtype="S6")
+        ass[()] = scalar
+        expected = scalar.astype("S")[:6]
+        assert cast[()] == expected
+        assert ass[()] == expected
+
+class TestNested:
+    def test_nested_simple(self):
+        initial = [1.2]
+        nested = initial
+        for i in range(np.MAXDIMS - 1):
+            nested = [nested]
+
+        arr = np.array(nested, dtype="float64")
+        assert arr.shape == (1,) * np.MAXDIMS
+        with pytest.raises(ValueError):
+            np.array([nested], dtype="float64")
+
+        with pytest.raises(ValueError, match=".*would exceed the maximum"):
+            np.array([nested])  # user must ask for `object` explicitly
+
+        arr = np.array([nested], dtype=object)
+        assert arr.dtype == np.dtype("O")
+        assert arr.shape == (1,) * np.MAXDIMS
+        assert arr.item() is initial
+
+    def test_pathological_self_containing(self):
+        # Test that this also works for two nested sequences
+        l = []
+        l.append(l)
+        arr = np.array([l, l, l], dtype=object)
+        assert arr.shape == (3,) + (1,) * (np.MAXDIMS - 1)
+
+        # Also check a ragged case:
+        arr = np.array([l, [None], l], dtype=object)
+        assert arr.shape == (3, 1)
+
+    @pytest.mark.parametrize("arraylike", arraylikes())
+    def test_nested_arraylikes(self, arraylike):
+        # We try storing an array like into an array, but the array-like
+        # will have too many dimensions.  This means the shape discovery
+        # decides that the array-like must be treated as an object (a special
+        # case of ragged discovery).  The result will be an array with one
+        # dimension less than the maximum dimensions, and the array being
+        # assigned to it (which does work for object or if `float(arraylike)`
+        # works).
+        initial = arraylike(np.ones((1, 1)))
+
+        nested = initial
+        for i in range(np.MAXDIMS - 1):
+            nested = [nested]
+
+        with pytest.raises(ValueError, match=".*would exceed the maximum"):
+            # It will refuse to assign the array into
+            np.array(nested, dtype="float64")
+
+        # If this is object, we end up assigning a (1, 1) array into (1,)
+        # (due to running out of dimensions), this is currently supported but
+        # a special case which is not ideal.
+        arr = np.array(nested, dtype=object)
+        assert arr.shape == (1,) * np.MAXDIMS
+        assert arr.item() == np.array(initial).item()
+
+    @pytest.mark.parametrize("arraylike", arraylikes())
+    def test_uneven_depth_ragged(self, arraylike):
+        arr = np.arange(4).reshape((2, 2))
+        arr = arraylike(arr)
+
+        # Array is ragged in the second dimension already:
+        out = np.array([arr, [arr]], dtype=object)
+        assert out.shape == (2,)
+        assert out[0] is arr
+        assert type(out[1]) is list
+
+        # Array is ragged in the third dimension:
+        with pytest.raises(ValueError):
+            # This is a broadcast error during assignment, because
+            # the array shape would be (2, 2, 2) but `arr[0, 0] = arr` fails.
+            np.array([arr, [arr, arr]], dtype=object)
+
+    def test_empty_sequence(self):
+        arr = np.array([[], [1], [[1]]], dtype=object)
+        assert arr.shape == (3,)
+
+        # The empty sequence stops further dimension discovery, so the
+        # result shape will be (0,) which leads to an error during:
+        with pytest.raises(ValueError):
+            np.array([[], np.empty((0, 1))], dtype=object)
+
+    def test_array_of_different_depths(self):
+        # When multiple arrays (or array-likes) are included in a
+        # sequences and have different depth, we currently discover
+        # as many dimensions as they share. (see also gh-17224)
+        arr = np.zeros((3, 2))
+        mismatch_first_dim = np.zeros((1, 2))
+        mismatch_second_dim = np.zeros((3, 3))
+
+        dtype, shape = _discover_array_parameters(
+            [arr, mismatch_second_dim], dtype=np.dtype("O"))
+        assert shape == (2, 3)
+
+        dtype, shape = _discover_array_parameters(
+            [arr, mismatch_first_dim], dtype=np.dtype("O"))
+        assert shape == (2,)
+        # The second case is currently supported because the arrays
+        # can be stored as objects:
+        res = np.asarray([arr, mismatch_first_dim], dtype=np.dtype("O"))
+        assert res[0] is arr
+        assert res[1] is mismatch_first_dim
+
+
+class TestBadSequences:
+    # These are tests for bad objects passed into `np.array`, in general
+    # these have undefined behaviour.  In the old code they partially worked
+    # when now they will fail.  We could (and maybe should) create a copy
+    # of all sequences to be safe against bad-actors.
+
+    def test_growing_list(self):
+        # List to coerce, `mylist` will append to it during coercion
+        obj = []
+        class mylist(list):
+            def __len__(self):
+                obj.append([1, 2])
+                return super().__len__()
+
+        obj.append(mylist([1, 2]))
+
+        with pytest.raises(RuntimeError):
+            np.array(obj)
+
+    # Note: We do not test a shrinking list.  These do very evil things
+    #       and the only way to fix them would be to copy all sequences.
+    #       (which may be a real option in the future).
+
+    def test_mutated_list(self):
+        # List to coerce, `mylist` will mutate the first element
+        obj = []
+        class mylist(list):
+            def __len__(self):
+                obj[0] = [2, 3]  # replace with a different list.
+                return super().__len__()
+
+        obj.append([2, 3])
+        obj.append(mylist([1, 2]))
+        # Does not crash:
+        np.array(obj)
+
+    def test_replace_0d_array(self):
+        # List to coerce, `mylist` will mutate the first element
+        obj = []
+        class baditem:
+            def __len__(self):
+                obj[0][0] = 2  # replace with a different list.
+                raise ValueError("not actually a sequence!")
+
+            def __getitem__(self):
+                pass
+
+        # Runs into a corner case in the new code, the `array(2)` is cached
+        # so replacing it invalidates the cache.
+        obj.append([np.array(2), baditem()])
+        with pytest.raises(RuntimeError):
+            np.array(obj)
+
+
+class TestArrayLikes:
+    @pytest.mark.parametrize("arraylike", arraylikes())
+    def test_0d_object_special_case(self, arraylike):
+        arr = np.array(0.)
+        obj = arraylike(arr)
+        # A single array-like is always converted:
+        res = np.array(obj, dtype=object)
+        assert_array_equal(arr, res)
+
+        # But a single 0-D nested array-like never:
+        res = np.array([obj], dtype=object)
+        assert res[0] is obj
+
+    @pytest.mark.parametrize("arraylike", arraylikes())
+    @pytest.mark.parametrize("arr", [np.array(0.), np.arange(4)])
+    def test_object_assignment_special_case(self, arraylike, arr):
+        obj = arraylike(arr)
+        empty = np.arange(1, dtype=object)
+        empty[:] = [obj]
+        assert empty[0] is obj
+
+    def test_0d_generic_special_case(self):
+        class ArraySubclass(np.ndarray):
+            def __float__(self):
+                raise TypeError("e.g. quantities raise on this")
+
+        arr = np.array(0.)
+        obj = arr.view(ArraySubclass)
+        res = np.array(obj)
+        # The subclass is simply cast:
+        assert_array_equal(arr, res)
+
+        # If the 0-D array-like is included, __float__ is currently
+        # guaranteed to be used.  We may want to change that, quantities
+        # and masked arrays half make use of this.
+        with pytest.raises(TypeError):
+            np.array([obj])
+
+        # The same holds for memoryview:
+        obj = memoryview(arr)
+        res = np.array(obj)
+        assert_array_equal(arr, res)
+        with pytest.raises(ValueError):
+            # The error type does not matter much here.
+            np.array([obj])
+
+    def test_arraylike_classes(self):
+        # The classes of array-likes should generally be acceptable to be
+        # stored inside a numpy (object) array.  This tests all of the
+        # special attributes (since all are checked during coercion).
+        arr = np.array(np.int64)
+        assert arr[()] is np.int64
+        arr = np.array([np.int64])
+        assert arr[0] is np.int64
+
+        # This also works for properties/unbound methods:
+        class ArrayLike:
+            @property
+            def __array_interface__(self):
+                pass
+
+            @property
+            def __array_struct__(self):
+                pass
+
+            def __array__(self):
+                pass
+
+        arr = np.array(ArrayLike)
+        assert arr[()] is ArrayLike
+        arr = np.array([ArrayLike])
+        assert arr[0] is ArrayLike
+
+    @pytest.mark.skipif(
+            np.dtype(np.intp).itemsize < 8, reason="Needs 64bit platform")
+    def test_too_large_array_error_paths(self):
+        """Test the error paths, including for memory leaks"""
+        arr = np.array(0, dtype="uint8")
+        # Guarantees that a contiguous copy won't work:
+        arr = np.broadcast_to(arr, 2**62)
+
+        for i in range(5):
+            # repeat, to ensure caching cannot have an effect:
+            with pytest.raises(MemoryError):
+                np.array(arr)
+            with pytest.raises(MemoryError):
+                np.array([arr])
+
+    @pytest.mark.parametrize("attribute",
+        ["__array_interface__", "__array__", "__array_struct__"])
+    @pytest.mark.parametrize("error", [RecursionError, MemoryError])
+    def test_bad_array_like_attributes(self, attribute, error):
+        # RecursionError and MemoryError are considered fatal. All errors
+        # (except AttributeError) should probably be raised in the future,
+        # but shapely made use of it, so it will require a deprecation.
+
+        class BadInterface:
+            def __getattr__(self, attr):
+                if attr == attribute:
+                    raise error
+                super().__getattr__(attr)
+
+        with pytest.raises(error):
+            np.array(BadInterface())
+
+    @pytest.mark.parametrize("error", [RecursionError, MemoryError])
+    def test_bad_array_like_bad_length(self, error):
+        # RecursionError and MemoryError are considered "critical" in
+        # sequences. We could expand this more generally though. (NumPy 1.20)
+        class BadSequence:
+            def __len__(self):
+                raise error
+            def __getitem__(self):
+                # must have getitem to be a Sequence
+                return 1
+
+        with pytest.raises(error):
+            np.array(BadSequence())
+
+
+class TestAsArray:
+    """Test expected behaviors of ``asarray``."""
+
+    def test_dtype_identity(self):
+        """Confirm the intended behavior for *dtype* kwarg.
+
+        The result of ``asarray()`` should have the dtype provided through the
+        keyword argument, when used. This forces unique array handles to be
+        produced for unique np.dtype objects, but (for equivalent dtypes), the
+        underlying data (the base object) is shared with the original array
+        object.
+
+        Ref https://github.com/numpy/numpy/issues/1468
+        """
+        int_array = np.array([1, 2, 3], dtype='i')
+        assert np.asarray(int_array) is int_array
+
+        # The character code resolves to the singleton dtype object provided
+        # by the numpy package.
+        assert np.asarray(int_array, dtype='i') is int_array
+
+        # Derive a dtype from n.dtype('i'), but add a metadata object to force
+        # the dtype to be distinct.
+        unequal_type = np.dtype('i', metadata={'spam': True})
+        annotated_int_array = np.asarray(int_array, dtype=unequal_type)
+        assert annotated_int_array is not int_array
+        assert annotated_int_array.base is int_array
+        # Create an equivalent descriptor with a new and distinct dtype
+        # instance.
+        equivalent_requirement = np.dtype('i', metadata={'spam': True})
+        annotated_int_array_alt = np.asarray(annotated_int_array,
+                                             dtype=equivalent_requirement)
+        assert unequal_type == equivalent_requirement
+        assert unequal_type is not equivalent_requirement
+        assert annotated_int_array_alt is not annotated_int_array
+        assert annotated_int_array_alt.dtype is equivalent_requirement
+
+        # Check the same logic for a pair of C types whose equivalence may vary
+        # between computing environments.
+        # Find an equivalent pair.
+        integer_type_codes = ('i', 'l', 'q')
+        integer_dtypes = [np.dtype(code) for code in integer_type_codes]
+        typeA = None
+        typeB = None
+        for typeA, typeB in permutations(integer_dtypes, r=2):
+            if typeA == typeB:
+                assert typeA is not typeB
+                break
+        assert isinstance(typeA, np.dtype) and isinstance(typeB, np.dtype)
+
+        # These ``asarray()`` calls may produce a new view or a copy,
+        # but never the same object.
+        long_int_array = np.asarray(int_array, dtype='l')
+        long_long_int_array = np.asarray(int_array, dtype='q')
+        assert long_int_array is not int_array
+        assert long_long_int_array is not int_array
+        assert np.asarray(long_int_array, dtype='q') is not long_int_array
+        array_a = np.asarray(int_array, dtype=typeA)
+        assert typeA == typeB
+        assert typeA is not typeB
+        assert array_a.dtype is typeA
+        assert array_a is not np.asarray(array_a, dtype=typeB)
+        assert np.asarray(array_a, dtype=typeB).dtype is typeB
+        assert array_a is np.asarray(array_a, dtype=typeB).base
+
+
+class TestSpecialAttributeLookupFailure:
+    # An exception was raised while fetching the attribute
+
+    class WeirdArrayLike:
+        @property
+        def __array__(self):
+            raise RuntimeError("oops!")
+
+    class WeirdArrayInterface:
+        @property
+        def __array_interface__(self):
+            raise RuntimeError("oops!")
+
+    def test_deprecated(self):
+        with pytest.raises(RuntimeError):
+            np.array(self.WeirdArrayLike())
+        with pytest.raises(RuntimeError):
+            np.array(self.WeirdArrayInterface())
+
+
+def test_subarray_from_array_construction():
+    # Arrays are more complex, since they "broadcast" on success:
+    arr = np.array([1, 2])
+
+    res = arr.astype("(2)i,")
+    assert_array_equal(res, [[1, 1], [2, 2]])
+
+    res = np.array(arr, dtype="(2)i,")
+
+    assert_array_equal(res, [[1, 1], [2, 2]])
+
+    res = np.array([[(1,), (2,)], arr], dtype="(2)i,")
+    assert_array_equal(res, [[[1, 1], [2, 2]], [[1, 1], [2, 2]]])
+
+    # Also try a multi-dimensional example:
+    arr = np.arange(5 * 2).reshape(5, 2)
+    expected = np.broadcast_to(arr[:, :, np.newaxis, np.newaxis], (5, 2, 2, 2))
+
+    res = arr.astype("(2,2)f")
+    assert_array_equal(res, expected)
+
+    res = np.array(arr, dtype="(2,2)f")
+    assert_array_equal(res, expected)
+
+
+def test_empty_string():
+    # Empty strings are unfortunately often converted to S1 and we need to
+    # make sure we are filling the S1 and not the (possibly) detected S0
+    # result.  This should likely just return S0 and if not maybe the decision
+    # to return S1 should be moved.
+    res = np.array([""] * 10, dtype="S")
+    assert_array_equal(res, np.array("\0", "S1"))
+    assert res.dtype == "S1"
+
+    arr = np.array([""] * 10, dtype=object)
+
+    res = arr.astype("S")
+    assert_array_equal(res, b"")
+    assert res.dtype == "S1"
+
+    res = np.array(arr, dtype="S")
+    assert_array_equal(res, b"")
+    # TODO: This is arguably weird/wrong, but seems old:
+    assert res.dtype == f"S{np.dtype('O').itemsize}"
+
+    res = np.array([[""] * 10, arr], dtype="S")
+    assert_array_equal(res, b"")
+    assert res.shape == (2, 10)
+    assert res.dtype == "S1"

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/core/tests/test_custom_dtypes.py

@@ -0,0 +1,253 @@
+import pytest
+
+import numpy as np
+from numpy.testing import assert_array_equal
+from numpy.core._multiarray_umath import (
+    _discover_array_parameters as discover_array_params, _get_sfloat_dtype)
+
+
+SF = _get_sfloat_dtype()
+
+
+class TestSFloat:
+    def _get_array(self, scaling, aligned=True):
+        if not aligned:
+            a = np.empty(3*8 + 1, dtype=np.uint8)[1:]
+            a = a.view(np.float64)
+            a[:] = [1., 2., 3.]
+        else:
+            a = np.array([1., 2., 3.])
+
+        a *= 1./scaling  # the casting code also uses the reciprocal.
+        return a.view(SF(scaling))
+
+    def test_sfloat_rescaled(self):
+        sf = SF(1.)
+        sf2 = sf.scaled_by(2.)
+        assert sf2.get_scaling() == 2.
+        sf6 = sf2.scaled_by(3.)
+        assert sf6.get_scaling() == 6.
+
+    def test_class_discovery(self):
+        # This does not test much, since we always discover the scaling as 1.
+        # But most of NumPy (when writing) does not understand DType classes
+        dt, _ = discover_array_params([1., 2., 3.], dtype=SF)
+        assert dt == SF(1.)
+
+    @pytest.mark.parametrize("scaling", [1., -1., 2.])
+    def test_scaled_float_from_floats(self, scaling):
+        a = np.array([1., 2., 3.], dtype=SF(scaling))
+
+        assert a.dtype.get_scaling() == scaling
+        assert_array_equal(scaling * a.view(np.float64), [1., 2., 3.])
+
+    def test_repr(self):
+        # Check the repr, mainly to cover the code paths:
+        assert repr(SF(scaling=1.)) == "_ScaledFloatTestDType(scaling=1.0)"
+
+    def test_dtype_name(self):
+        assert SF(1.).name == "_ScaledFloatTestDType64"
+
+    @pytest.mark.parametrize("scaling", [1., -1., 2.])
+    def test_sfloat_from_float(self, scaling):
+        a = np.array([1., 2., 3.]).astype(dtype=SF(scaling))
+
+        assert a.dtype.get_scaling() == scaling
+        assert_array_equal(scaling * a.view(np.float64), [1., 2., 3.])
+
+    @pytest.mark.parametrize("aligned", [True, False])
+    @pytest.mark.parametrize("scaling", [1., -1., 2.])
+    def test_sfloat_getitem(self, aligned, scaling):
+        a = self._get_array(1., aligned)
+        assert a.tolist() == [1., 2., 3.]
+
+    @pytest.mark.parametrize("aligned", [True, False])
+    def test_sfloat_casts(self, aligned):
+        a = self._get_array(1., aligned)
+
+        assert np.can_cast(a, SF(-1.), casting="equiv")
+        assert not np.can_cast(a, SF(-1.), casting="no")
+        na = a.astype(SF(-1.))
+        assert_array_equal(-1 * na.view(np.float64), a.view(np.float64))
+
+        assert np.can_cast(a, SF(2.), casting="same_kind")
+        assert not np.can_cast(a, SF(2.), casting="safe")
+        a2 = a.astype(SF(2.))
+        assert_array_equal(2 * a2.view(np.float64), a.view(np.float64))
+
+    @pytest.mark.parametrize("aligned", [True, False])
+    def test_sfloat_cast_internal_errors(self, aligned):
+        a = self._get_array(2e300, aligned)
+
+        with pytest.raises(TypeError,
+                match="error raised inside the core-loop: non-finite factor!"):
+            a.astype(SF(2e-300))
+
+    def test_sfloat_promotion(self):
+        assert np.result_type(SF(2.), SF(3.)) == SF(3.)
+        assert np.result_type(SF(3.), SF(2.)) == SF(3.)
+        # Float64 -> SF(1.) and then promotes normally, so both of this work:
+        assert np.result_type(SF(3.), np.float64) == SF(3.)
+        assert np.result_type(np.float64, SF(0.5)) == SF(1.)
+
+        # Test an undefined promotion:
+        with pytest.raises(TypeError):
+            np.result_type(SF(1.), np.int64)
+
+    def test_basic_multiply(self):
+        a = self._get_array(2.)
+        b = self._get_array(4.)
+
+        res = a * b
+        # multiplies dtype scaling and content separately:
+        assert res.dtype.get_scaling() == 8.
+        expected_view = a.view(np.float64) * b.view(np.float64)
+        assert_array_equal(res.view(np.float64), expected_view)
+
+    def test_possible_and_impossible_reduce(self):
+        # For reductions to work, the first and last operand must have the
+        # same dtype.  For this parametric DType that is not necessarily true.
+        a = self._get_array(2.)
+        # Addition reductin works (as of writing requires to pass initial
+        # because setting a scaled-float from the default `0` fails).
+        res = np.add.reduce(a, initial=0.)
+        assert res == a.astype(np.float64).sum()
+
+        # But each multiplication changes the factor, so a reduction is not
+        # possible (the relaxed version of the old refusal to handle any
+        # flexible dtype).
+        with pytest.raises(TypeError,
+                match="the resolved dtypes are not compatible"):
+            np.multiply.reduce(a)
+
+    def test_basic_ufunc_at(self):
+        float_a = np.array([1., 2., 3.])
+        b = self._get_array(2.)
+
+        float_b = b.view(np.float64).copy()
+        np.multiply.at(float_b, [1, 1, 1], float_a)
+        np.multiply.at(b, [1, 1, 1], float_a)
+
+        assert_array_equal(b.view(np.float64), float_b)
+
+    def test_basic_multiply_promotion(self):
+        float_a = np.array([1., 2., 3.])
+        b = self._get_array(2.)
+
+        res1 = float_a * b
+        res2 = b * float_a
+
+        # one factor is one, so we get the factor of b:
+        assert res1.dtype == res2.dtype == b.dtype
+        expected_view = float_a * b.view(np.float64)
+        assert_array_equal(res1.view(np.float64), expected_view)
+        assert_array_equal(res2.view(np.float64), expected_view)
+
+        # Check that promotion works when `out` is used:
+        np.multiply(b, float_a, out=res2)
+        with pytest.raises(TypeError):
+            # The promoter accepts this (maybe it should not), but the SFloat
+            # result cannot be cast to integer:
+            np.multiply(b, float_a, out=np.arange(3))
+
+    def test_basic_addition(self):
+        a = self._get_array(2.)
+        b = self._get_array(4.)
+
+        res = a + b
+        # addition uses the type promotion rules for the result:
+        assert res.dtype == np.result_type(a.dtype, b.dtype)
+        expected_view = (a.astype(res.dtype).view(np.float64) +
+                         b.astype(res.dtype).view(np.float64))
+        assert_array_equal(res.view(np.float64), expected_view)
+
+    def test_addition_cast_safety(self):
+        """The addition method is special for the scaled float, because it
+        includes the "cast" between different factors, thus cast-safety
+        is influenced by the implementation.
+        """
+        a = self._get_array(2.)
+        b = self._get_array(-2.)
+        c = self._get_array(3.)
+
+        # sign change is "equiv":
+        np.add(a, b, casting="equiv")
+        with pytest.raises(TypeError):
+            np.add(a, b, casting="no")
+
+        # Different factor is "same_kind" (default) so check that "safe" fails
+        with pytest.raises(TypeError):
+            np.add(a, c, casting="safe")
+
+        # Check that casting the output fails also (done by the ufunc here)
+        with pytest.raises(TypeError):
+            np.add(a, a, out=c, casting="safe")
+
+    @pytest.mark.parametrize("ufunc",
+            [np.logical_and, np.logical_or, np.logical_xor])
+    def test_logical_ufuncs_casts_to_bool(self, ufunc):
+        a = self._get_array(2.)
+        a[0] = 0.  # make sure first element is considered False.
+
+        float_equiv = a.astype(float)
+        expected = ufunc(float_equiv, float_equiv)
+        res = ufunc(a, a)
+        assert_array_equal(res, expected)
+
+        # also check that the same works for reductions:
+        expected = ufunc.reduce(float_equiv)
+        res = ufunc.reduce(a)
+        assert_array_equal(res, expected)
+
+        # The output casting does not match the bool, bool -> bool loop:
+        with pytest.raises(TypeError):
+            ufunc(a, a, out=np.empty(a.shape, dtype=int), casting="equiv")
+
+    def test_wrapped_and_wrapped_reductions(self):
+        a = self._get_array(2.)
+        float_equiv = a.astype(float)
+
+        expected = np.hypot(float_equiv, float_equiv)
+        res = np.hypot(a, a)
+        assert res.dtype == a.dtype
+        res_float = res.view(np.float64) * 2
+        assert_array_equal(res_float, expected)
+
+        # Also check reduction (keepdims, due to incorrect getitem)
+        res = np.hypot.reduce(a, keepdims=True)
+        assert res.dtype == a.dtype
+        expected = np.hypot.reduce(float_equiv, keepdims=True)
+        assert res.view(np.float64) * 2 == expected
+
+    def test_astype_class(self):
+        # Very simple test that we accept `.astype()` also on the class.
+        # ScaledFloat always returns the default descriptor, but it does
+        # check the relevant code paths.
+        arr = np.array([1., 2., 3.], dtype=object)
+
+        res = arr.astype(SF)  # passing the class class
+        expected = arr.astype(SF(1.))  # above will have discovered 1. scaling
+        assert_array_equal(res.view(np.float64), expected.view(np.float64))
+
+    def test_creation_class(self):
+        arr1 = np.array([1., 2., 3.], dtype=SF)
+        assert arr1.dtype == SF(1.)
+        arr2 = np.array([1., 2., 3.], dtype=SF(1.))
+        assert_array_equal(arr1.view(np.float64), arr2.view(np.float64))
+
+
+def test_type_pickle():
+    # can't actually unpickle, but we can pickle (if in namespace)
+    import pickle
+
+    np._ScaledFloatTestDType = SF
+
+    s = pickle.dumps(SF)
+    res = pickle.loads(s)
+    assert res is SF
+
+    del np._ScaledFloatTestDType
+
+
+def test_is_numeric():
+    assert SF._is_numeric

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/core/tests/test_function_base.py

@@ -0,0 +1,446 @@
+import pytest
+from numpy import (
+    logspace, linspace, geomspace, dtype, array, sctypes, arange, isnan,
+    ndarray, sqrt, nextafter, stack, errstate
+    )
+from numpy.testing import (
+    assert_, assert_equal, assert_raises, assert_array_equal, assert_allclose,
+    )
+
+
+class PhysicalQuantity(float):
+    def __new__(cls, value):
+        return float.__new__(cls, value)
+
+    def __add__(self, x):
+        assert_(isinstance(x, PhysicalQuantity))
+        return PhysicalQuantity(float(x) + float(self))
+    __radd__ = __add__
+
+    def __sub__(self, x):
+        assert_(isinstance(x, PhysicalQuantity))
+        return PhysicalQuantity(float(self) - float(x))
+
+    def __rsub__(self, x):
+        assert_(isinstance(x, PhysicalQuantity))
+        return PhysicalQuantity(float(x) - float(self))
+
+    def __mul__(self, x):
+        return PhysicalQuantity(float(x) * float(self))
+    __rmul__ = __mul__
+
+    def __div__(self, x):
+        return PhysicalQuantity(float(self) / float(x))
+
+    def __rdiv__(self, x):
+        return PhysicalQuantity(float(x) / float(self))
+
+
+class PhysicalQuantity2(ndarray):
+    __array_priority__ = 10
+
+
+class TestLogspace:
+
+    def test_basic(self):
+        y = logspace(0, 6)
+        assert_(len(y) == 50)
+        y = logspace(0, 6, num=100)
+        assert_(y[-1] == 10 ** 6)
+        y = logspace(0, 6, endpoint=False)
+        assert_(y[-1] < 10 ** 6)
+        y = logspace(0, 6, num=7)
+        assert_array_equal(y, [1, 10, 100, 1e3, 1e4, 1e5, 1e6])
+
+    def test_start_stop_array(self):
+        start = array([0., 1.])
+        stop = array([6., 7.])
+        t1 = logspace(start, stop, 6)
+        t2 = stack([logspace(_start, _stop, 6)
+                    for _start, _stop in zip(start, stop)], axis=1)
+        assert_equal(t1, t2)
+        t3 = logspace(start, stop[0], 6)
+        t4 = stack([logspace(_start, stop[0], 6)
+                    for _start in start], axis=1)
+        assert_equal(t3, t4)
+        t5 = logspace(start, stop, 6, axis=-1)
+        assert_equal(t5, t2.T)
+
+    @pytest.mark.parametrize("axis", [0, 1, -1])
+    def test_base_array(self, axis: int):
+        start = 1
+        stop = 2
+        num = 6
+        base = array([1, 2])
+        t1 = logspace(start, stop, num=num, base=base, axis=axis)
+        t2 = stack(
+            [logspace(start, stop, num=num, base=_base) for _base in base],
+            axis=(axis + 1) % t1.ndim,
+        )
+        assert_equal(t1, t2)
+
+    @pytest.mark.parametrize("axis", [0, 1, -1])
+    def test_stop_base_array(self, axis: int):
+        start = 1
+        stop = array([2, 3])
+        num = 6
+        base = array([1, 2])
+        t1 = logspace(start, stop, num=num, base=base, axis=axis)
+        t2 = stack(
+            [logspace(start, _stop, num=num, base=_base)
+             for _stop, _base in zip(stop, base)],
+            axis=(axis + 1) % t1.ndim,
+        )
+        assert_equal(t1, t2)
+
+    def test_dtype(self):
+        y = logspace(0, 6, dtype='float32')
+        assert_equal(y.dtype, dtype('float32'))
+        y = logspace(0, 6, dtype='float64')
+        assert_equal(y.dtype, dtype('float64'))
+        y = logspace(0, 6, dtype='int32')
+        assert_equal(y.dtype, dtype('int32'))
+
+    def test_physical_quantities(self):
+        a = PhysicalQuantity(1.0)
+        b = PhysicalQuantity(5.0)
+        assert_equal(logspace(a, b), logspace(1.0, 5.0))
+
+    def test_subclass(self):
+        a = array(1).view(PhysicalQuantity2)
+        b = array(7).view(PhysicalQuantity2)
+        ls = logspace(a, b)
+        assert type(ls) is PhysicalQuantity2
+        assert_equal(ls, logspace(1.0, 7.0))
+        ls = logspace(a, b, 1)
+        assert type(ls) is PhysicalQuantity2
+        assert_equal(ls, logspace(1.0, 7.0, 1))
+
+
+class TestGeomspace:
+
+    def test_basic(self):
+        y = geomspace(1, 1e6)
+        assert_(len(y) == 50)
+        y = geomspace(1, 1e6, num=100)
+        assert_(y[-1] == 10 ** 6)
+        y = geomspace(1, 1e6, endpoint=False)
+        assert_(y[-1] < 10 ** 6)
+        y = geomspace(1, 1e6, num=7)
+        assert_array_equal(y, [1, 10, 100, 1e3, 1e4, 1e5, 1e6])
+
+        y = geomspace(8, 2, num=3)
+        assert_allclose(y, [8, 4, 2])
+        assert_array_equal(y.imag, 0)
+
+        y = geomspace(-1, -100, num=3)
+        assert_array_equal(y, [-1, -10, -100])
+        assert_array_equal(y.imag, 0)
+
+        y = geomspace(-100, -1, num=3)
+        assert_array_equal(y, [-100, -10, -1])
+        assert_array_equal(y.imag, 0)
+
+    def test_boundaries_match_start_and_stop_exactly(self):
+        # make sure that the boundaries of the returned array exactly
+        # equal 'start' and 'stop' - this isn't obvious because
+        # np.exp(np.log(x)) isn't necessarily exactly equal to x
+        start = 0.3
+        stop = 20.3
+
+        y = geomspace(start, stop, num=1)
+        assert_equal(y[0], start)
+
+        y = geomspace(start, stop, num=1, endpoint=False)
+        assert_equal(y[0], start)
+
+        y = geomspace(start, stop, num=3)
+        assert_equal(y[0], start)
+        assert_equal(y[-1], stop)
+
+        y = geomspace(start, stop, num=3, endpoint=False)
+        assert_equal(y[0], start)
+
+    def test_nan_interior(self):
+        with errstate(invalid='ignore'):
+            y = geomspace(-3, 3, num=4)
+
+        assert_equal(y[0], -3.0)
+        assert_(isnan(y[1:-1]).all())
+        assert_equal(y[3], 3.0)
+
+        with errstate(invalid='ignore'):
+            y = geomspace(-3, 3, num=4, endpoint=False)
+
+        assert_equal(y[0], -3.0)
+        assert_(isnan(y[1:]).all())
+
+    def test_complex(self):
+        # Purely imaginary
+        y = geomspace(1j, 16j, num=5)
+        assert_allclose(y, [1j, 2j, 4j, 8j, 16j])
+        assert_array_equal(y.real, 0)
+
+        y = geomspace(-4j, -324j, num=5)
+        assert_allclose(y, [-4j, -12j, -36j, -108j, -324j])
+        assert_array_equal(y.real, 0)
+
+        y = geomspace(1+1j, 1000+1000j, num=4)
+        assert_allclose(y, [1+1j, 10+10j, 100+100j, 1000+1000j])
+
+        y = geomspace(-1+1j, -1000+1000j, num=4)
+        assert_allclose(y, [-1+1j, -10+10j, -100+100j, -1000+1000j])
+
+        # Logarithmic spirals
+        y = geomspace(-1, 1, num=3, dtype=complex)
+        assert_allclose(y, [-1, 1j, +1])
+
+        y = geomspace(0+3j, -3+0j, 3)
+        assert_allclose(y, [0+3j, -3/sqrt(2)+3j/sqrt(2), -3+0j])
+        y = geomspace(0+3j, 3+0j, 3)
+        assert_allclose(y, [0+3j, 3/sqrt(2)+3j/sqrt(2), 3+0j])
+        y = geomspace(-3+0j, 0-3j, 3)
+        assert_allclose(y, [-3+0j, -3/sqrt(2)-3j/sqrt(2), 0-3j])
+        y = geomspace(0+3j, -3+0j, 3)
+        assert_allclose(y, [0+3j, -3/sqrt(2)+3j/sqrt(2), -3+0j])
+        y = geomspace(-2-3j, 5+7j, 7)
+        assert_allclose(y, [-2-3j, -0.29058977-4.15771027j,
+                            2.08885354-4.34146838j, 4.58345529-3.16355218j,
+                            6.41401745-0.55233457j, 6.75707386+3.11795092j,
+                            5+7j])
+
+        # Type promotion should prevent the -5 from becoming a NaN
+        y = geomspace(3j, -5, 2)
+        assert_allclose(y, [3j, -5])
+        y = geomspace(-5, 3j, 2)
+        assert_allclose(y, [-5, 3j])
+
+    def test_dtype(self):
+        y = geomspace(1, 1e6, dtype='float32')
+        assert_equal(y.dtype, dtype('float32'))
+        y = geomspace(1, 1e6, dtype='float64')
+        assert_equal(y.dtype, dtype('float64'))
+        y = geomspace(1, 1e6, dtype='int32')
+        assert_equal(y.dtype, dtype('int32'))
+
+        # Native types
+        y = geomspace(1, 1e6, dtype=float)
+        assert_equal(y.dtype, dtype('float_'))
+        y = geomspace(1, 1e6, dtype=complex)
+        assert_equal(y.dtype, dtype('complex'))
+
+    def test_start_stop_array_scalar(self):
+        lim1 = array([120, 100], dtype="int8")
+        lim2 = array([-120, -100], dtype="int8")
+        lim3 = array([1200, 1000], dtype="uint16")
+        t1 = geomspace(lim1[0], lim1[1], 5)
+        t2 = geomspace(lim2[0], lim2[1], 5)
+        t3 = geomspace(lim3[0], lim3[1], 5)
+        t4 = geomspace(120.0, 100.0, 5)
+        t5 = geomspace(-120.0, -100.0, 5)
+        t6 = geomspace(1200.0, 1000.0, 5)
+
+        # t3 uses float32, t6 uses float64
+        assert_allclose(t1, t4, rtol=1e-2)
+        assert_allclose(t2, t5, rtol=1e-2)
+        assert_allclose(t3, t6, rtol=1e-5)
+
+    def test_start_stop_array(self):
+        # Try to use all special cases.
+        start = array([1.e0, 32., 1j, -4j, 1+1j, -1])
+        stop = array([1.e4, 2., 16j, -324j, 10000+10000j, 1])
+        t1 = geomspace(start, stop, 5)
+        t2 = stack([geomspace(_start, _stop, 5)
+                    for _start, _stop in zip(start, stop)], axis=1)
+        assert_equal(t1, t2)
+        t3 = geomspace(start, stop[0], 5)
+        t4 = stack([geomspace(_start, stop[0], 5)
+                    for _start in start], axis=1)
+        assert_equal(t3, t4)
+        t5 = geomspace(start, stop, 5, axis=-1)
+        assert_equal(t5, t2.T)
+
+    def test_physical_quantities(self):
+        a = PhysicalQuantity(1.0)
+        b = PhysicalQuantity(5.0)
+        assert_equal(geomspace(a, b), geomspace(1.0, 5.0))
+
+    def test_subclass(self):
+        a = array(1).view(PhysicalQuantity2)
+        b = array(7).view(PhysicalQuantity2)
+        gs = geomspace(a, b)
+        assert type(gs) is PhysicalQuantity2
+        assert_equal(gs, geomspace(1.0, 7.0))
+        gs = geomspace(a, b, 1)
+        assert type(gs) is PhysicalQuantity2
+        assert_equal(gs, geomspace(1.0, 7.0, 1))
+
+    def test_bounds(self):
+        assert_raises(ValueError, geomspace, 0, 10)
+        assert_raises(ValueError, geomspace, 10, 0)
+        assert_raises(ValueError, geomspace, 0, 0)
+
+
+class TestLinspace:
+
+    def test_basic(self):
+        y = linspace(0, 10)
+        assert_(len(y) == 50)
+        y = linspace(2, 10, num=100)
+        assert_(y[-1] == 10)
+        y = linspace(2, 10, endpoint=False)
+        assert_(y[-1] < 10)
+        assert_raises(ValueError, linspace, 0, 10, num=-1)
+
+    def test_corner(self):
+        y = list(linspace(0, 1, 1))
+        assert_(y == [0.0], y)
+        assert_raises(TypeError, linspace, 0, 1, num=2.5)
+
+    def test_type(self):
+        t1 = linspace(0, 1, 0).dtype
+        t2 = linspace(0, 1, 1).dtype
+        t3 = linspace(0, 1, 2).dtype
+        assert_equal(t1, t2)
+        assert_equal(t2, t3)
+
+    def test_dtype(self):
+        y = linspace(0, 6, dtype='float32')
+        assert_equal(y.dtype, dtype('float32'))
+        y = linspace(0, 6, dtype='float64')
+        assert_equal(y.dtype, dtype('float64'))
+        y = linspace(0, 6, dtype='int32')
+        assert_equal(y.dtype, dtype('int32'))
+
+    def test_start_stop_array_scalar(self):
+        lim1 = array([-120, 100], dtype="int8")
+        lim2 = array([120, -100], dtype="int8")
+        lim3 = array([1200, 1000], dtype="uint16")
+        t1 = linspace(lim1[0], lim1[1], 5)
+        t2 = linspace(lim2[0], lim2[1], 5)
+        t3 = linspace(lim3[0], lim3[1], 5)
+        t4 = linspace(-120.0, 100.0, 5)
+        t5 = linspace(120.0, -100.0, 5)
+        t6 = linspace(1200.0, 1000.0, 5)
+        assert_equal(t1, t4)
+        assert_equal(t2, t5)
+        assert_equal(t3, t6)
+
+    def test_start_stop_array(self):
+        start = array([-120, 120], dtype="int8")
+        stop = array([100, -100], dtype="int8")
+        t1 = linspace(start, stop, 5)
+        t2 = stack([linspace(_start, _stop, 5)
+                    for _start, _stop in zip(start, stop)], axis=1)
+        assert_equal(t1, t2)
+        t3 = linspace(start, stop[0], 5)
+        t4 = stack([linspace(_start, stop[0], 5)
+                    for _start in start], axis=1)
+        assert_equal(t3, t4)
+        t5 = linspace(start, stop, 5, axis=-1)
+        assert_equal(t5, t2.T)
+
+    def test_complex(self):
+        lim1 = linspace(1 + 2j, 3 + 4j, 5)
+        t1 = array([1.0+2.j, 1.5+2.5j,  2.0+3j, 2.5+3.5j, 3.0+4j])
+        lim2 = linspace(1j, 10, 5)
+        t2 = array([0.0+1.j, 2.5+0.75j, 5.0+0.5j, 7.5+0.25j, 10.0+0j])
+        assert_equal(lim1, t1)
+        assert_equal(lim2, t2)
+
+    def test_physical_quantities(self):
+        a = PhysicalQuantity(0.0)
+        b = PhysicalQuantity(1.0)
+        assert_equal(linspace(a, b), linspace(0.0, 1.0))
+
+    def test_subclass(self):
+        a = array(0).view(PhysicalQuantity2)
+        b = array(1).view(PhysicalQuantity2)
+        ls = linspace(a, b)
+        assert type(ls) is PhysicalQuantity2
+        assert_equal(ls, linspace(0.0, 1.0))
+        ls = linspace(a, b, 1)
+        assert type(ls) is PhysicalQuantity2
+        assert_equal(ls, linspace(0.0, 1.0, 1))
+
+    def test_array_interface(self):
+        # Regression test for https://github.com/numpy/numpy/pull/6659
+        # Ensure that start/stop can be objects that implement
+        # __array_interface__ and are convertible to numeric scalars
+
+        class Arrayish:
+            """
+            A generic object that supports the __array_interface__ and hence
+            can in principle be converted to a numeric scalar, but is not
+            otherwise recognized as numeric, but also happens to support
+            multiplication by floats.
+
+            Data should be an object that implements the buffer interface,
+            and contains at least 4 bytes.
+            """
+
+            def __init__(self, data):
+                self._data = data
+
+            @property
+            def __array_interface__(self):
+                return {'shape': (), 'typestr': '<i4', 'data': self._data,
+                        'version': 3}
+
+            def __mul__(self, other):
+                # For the purposes of this test any multiplication is an
+                # identity operation :)
+                return self
+
+        one = Arrayish(array(1, dtype='<i4'))
+        five = Arrayish(array(5, dtype='<i4'))
+
+        assert_equal(linspace(one, five), linspace(1, 5))
+
+    def test_denormal_numbers(self):
+        # Regression test for gh-5437. Will probably fail when compiled
+        # with ICC, which flushes denormals to zero
+        for ftype in sctypes['float']:
+            stop = nextafter(ftype(0), ftype(1)) * 5  # A denormal number
+            assert_(any(linspace(0, stop, 10, endpoint=False, dtype=ftype)))
+
+    def test_equivalent_to_arange(self):
+        for j in range(1000):
+            assert_equal(linspace(0, j, j+1, dtype=int),
+                         arange(j+1, dtype=int))
+
+    def test_retstep(self):
+        for num in [0, 1, 2]:
+            for ept in [False, True]:
+                y = linspace(0, 1, num, endpoint=ept, retstep=True)
+                assert isinstance(y, tuple) and len(y) == 2
+                if num == 2:
+                    y0_expect = [0.0, 1.0] if ept else [0.0, 0.5]
+                    assert_array_equal(y[0], y0_expect)
+                    assert_equal(y[1], y0_expect[1])
+                elif num == 1 and not ept:
+                    assert_array_equal(y[0], [0.0])
+                    assert_equal(y[1], 1.0)
+                else:
+                    assert_array_equal(y[0], [0.0][:num])
+                    assert isnan(y[1])
+
+    def test_object(self):
+        start = array(1, dtype='O')
+        stop = array(2, dtype='O')
+        y = linspace(start, stop, 3)
+        assert_array_equal(y, array([1., 1.5, 2.]))
+                    
+    def test_round_negative(self):
+        y = linspace(-1, 3, num=8, dtype=int)
+        t = array([-1, -1, 0, 0, 1, 1, 2, 3], dtype=int)
+        assert_array_equal(y, t)
+
+    def test_any_step_zero_and_not_mult_inplace(self):
+        # any_step_zero is True, _mult_inplace is False
+        start = array([0.0, 1.0])
+        stop = array([2.0, 1.0])
+        y = linspace(start, stop, 3)
+        assert_array_equal(y, array([[0.0, 1.0], [1.0, 1.0], [2.0, 1.0]]))
+    
+

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/core/tests/test_getlimits.py

@@ -0,0 +1,194 @@
+""" Test functions for limits module.
+
+"""
+import warnings
+import numpy as np
+import pytest
+from numpy.core import finfo, iinfo
+from numpy import half, single, double, longdouble
+from numpy.testing import assert_equal, assert_, assert_raises
+from numpy.core.getlimits import _discovered_machar, _float_ma
+
+##################################################
+
+class TestPythonFloat:
+    def test_singleton(self):
+        ftype = finfo(float)
+        ftype2 = finfo(float)
+        assert_equal(id(ftype), id(ftype2))
+
+class TestHalf:
+    def test_singleton(self):
+        ftype = finfo(half)
+        ftype2 = finfo(half)
+        assert_equal(id(ftype), id(ftype2))
+
+class TestSingle:
+    def test_singleton(self):
+        ftype = finfo(single)
+        ftype2 = finfo(single)
+        assert_equal(id(ftype), id(ftype2))
+
+class TestDouble:
+    def test_singleton(self):
+        ftype = finfo(double)
+        ftype2 = finfo(double)
+        assert_equal(id(ftype), id(ftype2))
+
+class TestLongdouble:
+    def test_singleton(self):
+        ftype = finfo(longdouble)
+        ftype2 = finfo(longdouble)
+        assert_equal(id(ftype), id(ftype2))
+
+def assert_finfo_equal(f1, f2):
+    # assert two finfo instances have the same attributes
+    for attr in ('bits', 'eps', 'epsneg', 'iexp', 'machep',
+                 'max', 'maxexp', 'min', 'minexp', 'negep', 'nexp',
+                 'nmant', 'precision', 'resolution', 'tiny',
+                 'smallest_normal', 'smallest_subnormal'):
+        assert_equal(getattr(f1, attr), getattr(f2, attr),
+                     f'finfo instances {f1} and {f2} differ on {attr}')
+
+def assert_iinfo_equal(i1, i2):
+    # assert two iinfo instances have the same attributes
+    for attr in ('bits', 'min', 'max'):
+        assert_equal(getattr(i1, attr), getattr(i2, attr),
+                     f'iinfo instances {i1} and {i2} differ on {attr}')
+
+class TestFinfo:
+    def test_basic(self):
+        dts = list(zip(['f2', 'f4', 'f8', 'c8', 'c16'],
+                       [np.float16, np.float32, np.float64, np.complex64,
+                        np.complex128]))
+        for dt1, dt2 in dts:
+            assert_finfo_equal(finfo(dt1), finfo(dt2))
+
+        assert_raises(ValueError, finfo, 'i4')
+
+    def test_regression_gh23108(self):
+        # np.float32(1.0) and np.float64(1.0) have the same hash and are
+        # equal under the == operator
+        f1 = np.finfo(np.float32(1.0))
+        f2 = np.finfo(np.float64(1.0))
+        assert f1 != f2
+
+    def test_regression_gh23867(self):
+        class NonHashableWithDtype:
+            __hash__ = None
+            dtype = np.dtype('float32')
+  
+        x = NonHashableWithDtype()
+        assert np.finfo(x) == np.finfo(x.dtype)
+        
+
+class TestIinfo:
+    def test_basic(self):
+        dts = list(zip(['i1', 'i2', 'i4', 'i8',
+                   'u1', 'u2', 'u4', 'u8'],
+                  [np.int8, np.int16, np.int32, np.int64,
+                   np.uint8, np.uint16, np.uint32, np.uint64]))
+        for dt1, dt2 in dts:
+            assert_iinfo_equal(iinfo(dt1), iinfo(dt2))
+
+        assert_raises(ValueError, iinfo, 'f4')
+
+    def test_unsigned_max(self):
+        types = np.sctypes['uint']
+        for T in types:
+            with np.errstate(over="ignore"):
+                max_calculated = T(0) - T(1)
+            assert_equal(iinfo(T).max, max_calculated)
+
+class TestRepr:
+    def test_iinfo_repr(self):
+        expected = "iinfo(min=-32768, max=32767, dtype=int16)"
+        assert_equal(repr(np.iinfo(np.int16)), expected)
+
+    def test_finfo_repr(self):
+        expected = "finfo(resolution=1e-06, min=-3.4028235e+38," + \
+                   " max=3.4028235e+38, dtype=float32)"
+        assert_equal(repr(np.finfo(np.float32)), expected)
+
+
+def test_instances():
+    # Test the finfo and iinfo results on numeric instances agree with
+    # the results on the corresponding types
+
+    for c in [int, np.int16, np.int32, np.int64]:
+        class_iinfo = iinfo(c)
+        instance_iinfo = iinfo(c(12))
+
+        assert_iinfo_equal(class_iinfo, instance_iinfo)
+
+    for c in [float, np.float16, np.float32, np.float64]:
+        class_finfo = finfo(c)
+        instance_finfo = finfo(c(1.2))
+        assert_finfo_equal(class_finfo, instance_finfo)
+
+    with pytest.raises(ValueError):
+        iinfo(10.)
+
+    with pytest.raises(ValueError):
+        iinfo('hi')
+
+    with pytest.raises(ValueError):
+        finfo(np.int64(1))
+
+
+def assert_ma_equal(discovered, ma_like):
+    # Check MachAr-like objects same as calculated MachAr instances
+    for key, value in discovered.__dict__.items():
+        assert_equal(value, getattr(ma_like, key))
+        if hasattr(value, 'shape'):
+            assert_equal(value.shape, getattr(ma_like, key).shape)
+            assert_equal(value.dtype, getattr(ma_like, key).dtype)
+
+
+def test_known_types():
+    # Test we are correctly compiling parameters for known types
+    for ftype, ma_like in ((np.float16, _float_ma[16]),
+                           (np.float32, _float_ma[32]),
+                           (np.float64, _float_ma[64])):
+        assert_ma_equal(_discovered_machar(ftype), ma_like)
+    # Suppress warning for broken discovery of double double on PPC
+    with np.errstate(all='ignore'):
+        ld_ma = _discovered_machar(np.longdouble)
+    bytes = np.dtype(np.longdouble).itemsize
+    if (ld_ma.it, ld_ma.maxexp) == (63, 16384) and bytes in (12, 16):
+        # 80-bit extended precision
+        assert_ma_equal(ld_ma, _float_ma[80])
+    elif (ld_ma.it, ld_ma.maxexp) == (112, 16384) and bytes == 16:
+        # IEE 754 128-bit
+        assert_ma_equal(ld_ma, _float_ma[128])
+
+
+def test_subnormal_warning():
+    """Test that the subnormal is zero warning is not being raised."""
+    with np.errstate(all='ignore'):
+        ld_ma = _discovered_machar(np.longdouble)
+    bytes = np.dtype(np.longdouble).itemsize
+    with warnings.catch_warnings(record=True) as w:
+        warnings.simplefilter('always')
+        if (ld_ma.it, ld_ma.maxexp) == (63, 16384) and bytes in (12, 16):
+            # 80-bit extended precision
+            ld_ma.smallest_subnormal
+            assert len(w) == 0
+        elif (ld_ma.it, ld_ma.maxexp) == (112, 16384) and bytes == 16:
+            # IEE 754 128-bit
+            ld_ma.smallest_subnormal
+            assert len(w) == 0
+        else:
+            # Double double
+            ld_ma.smallest_subnormal
+            # This test may fail on some platforms
+            assert len(w) == 0
+
+
+def test_plausible_finfo():
+    # Assert that finfo returns reasonable results for all types
+    for ftype in np.sctypes['float'] + np.sctypes['complex']:
+        info = np.finfo(ftype)
+        assert_(info.nmant > 1)
+        assert_(info.minexp < -1)
+        assert_(info.maxexp > 1)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/core/tests/test_hashtable.py

@@ -0,0 +1,30 @@
+import pytest
+
+import random
+from numpy.core._multiarray_tests import identityhash_tester
+
+
+@pytest.mark.parametrize("key_length", [1, 3, 6])
+@pytest.mark.parametrize("length", [1, 16, 2000])
+def test_identity_hashtable(key_length, length):
+    # use a 30 object pool for everything (duplicates will happen)
+    pool = [object() for i in range(20)]
+    keys_vals = []
+    for i in range(length):
+        keys = tuple(random.choices(pool, k=key_length))
+        keys_vals.append((keys, random.choice(pool)))
+
+    dictionary = dict(keys_vals)
+
+    # add a random item at the end:
+    keys_vals.append(random.choice(keys_vals))
+    # the expected one could be different with duplicates:
+    expected = dictionary[keys_vals[-1][0]]
+
+    res = identityhash_tester(key_length, keys_vals, replace=True)
+    assert res is expected
+
+    # check that ensuring one duplicate definitely raises:
+    keys_vals.insert(0, keys_vals[-2])
+    with pytest.raises(RuntimeError):
+        identityhash_tester(key_length, keys_vals)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/core/tests/test_indexing.py

@@ -0,0 +1,1417 @@
+import sys
+import warnings
+import functools
+import operator
+
+import pytest
+
+import numpy as np
+from numpy.core._multiarray_tests import array_indexing
+from itertools import product
+from numpy.testing import (
+    assert_, assert_equal, assert_raises, assert_raises_regex,
+    assert_array_equal, assert_warns, HAS_REFCOUNT, IS_WASM
+    )
+
+
+class TestIndexing:
+    def test_index_no_floats(self):
+        a = np.array([[[5]]])
+
+        assert_raises(IndexError, lambda: a[0.0])
+        assert_raises(IndexError, lambda: a[0, 0.0])
+        assert_raises(IndexError, lambda: a[0.0, 0])
+        assert_raises(IndexError, lambda: a[0.0,:])
+        assert_raises(IndexError, lambda: a[:, 0.0])
+        assert_raises(IndexError, lambda: a[:, 0.0,:])
+        assert_raises(IndexError, lambda: a[0.0,:,:])
+        assert_raises(IndexError, lambda: a[0, 0, 0.0])
+        assert_raises(IndexError, lambda: a[0.0, 0, 0])
+        assert_raises(IndexError, lambda: a[0, 0.0, 0])
+        assert_raises(IndexError, lambda: a[-1.4])
+        assert_raises(IndexError, lambda: a[0, -1.4])
+        assert_raises(IndexError, lambda: a[-1.4, 0])
+        assert_raises(IndexError, lambda: a[-1.4,:])
+        assert_raises(IndexError, lambda: a[:, -1.4])
+        assert_raises(IndexError, lambda: a[:, -1.4,:])
+        assert_raises(IndexError, lambda: a[-1.4,:,:])
+        assert_raises(IndexError, lambda: a[0, 0, -1.4])
+        assert_raises(IndexError, lambda: a[-1.4, 0, 0])
+        assert_raises(IndexError, lambda: a[0, -1.4, 0])
+        assert_raises(IndexError, lambda: a[0.0:, 0.0])
+        assert_raises(IndexError, lambda: a[0.0:, 0.0,:])
+
+    def test_slicing_no_floats(self):
+        a = np.array([[5]])
+
+        # start as float.
+        assert_raises(TypeError, lambda: a[0.0:])
+        assert_raises(TypeError, lambda: a[0:, 0.0:2])
+        assert_raises(TypeError, lambda: a[0.0::2, :0])
+        assert_raises(TypeError, lambda: a[0.0:1:2,:])
+        assert_raises(TypeError, lambda: a[:, 0.0:])
+        # stop as float.
+        assert_raises(TypeError, lambda: a[:0.0])
+        assert_raises(TypeError, lambda: a[:0, 1:2.0])
+        assert_raises(TypeError, lambda: a[:0.0:2, :0])
+        assert_raises(TypeError, lambda: a[:0.0,:])
+        assert_raises(TypeError, lambda: a[:, 0:4.0:2])
+        # step as float.
+        assert_raises(TypeError, lambda: a[::1.0])
+        assert_raises(TypeError, lambda: a[0:, :2:2.0])
+        assert_raises(TypeError, lambda: a[1::4.0, :0])
+        assert_raises(TypeError, lambda: a[::5.0,:])
+        assert_raises(TypeError, lambda: a[:, 0:4:2.0])
+        # mixed.
+        assert_raises(TypeError, lambda: a[1.0:2:2.0])
+        assert_raises(TypeError, lambda: a[1.0::2.0])
+        assert_raises(TypeError, lambda: a[0:, :2.0:2.0])
+        assert_raises(TypeError, lambda: a[1.0:1:4.0, :0])
+        assert_raises(TypeError, lambda: a[1.0:5.0:5.0,:])
+        assert_raises(TypeError, lambda: a[:, 0.4:4.0:2.0])
+        # should still get the DeprecationWarning if step = 0.
+        assert_raises(TypeError, lambda: a[::0.0])
+
+    def test_index_no_array_to_index(self):
+        # No non-scalar arrays.
+        a = np.array([[[1]]])
+
+        assert_raises(TypeError, lambda: a[a:a:a])
+
+    def test_none_index(self):
+        # `None` index adds newaxis
+        a = np.array([1, 2, 3])
+        assert_equal(a[None], a[np.newaxis])
+        assert_equal(a[None].ndim, a.ndim + 1)
+
+    def test_empty_tuple_index(self):
+        # Empty tuple index creates a view
+        a = np.array([1, 2, 3])
+        assert_equal(a[()], a)
+        assert_(a[()].base is a)
+        a = np.array(0)
+        assert_(isinstance(a[()], np.int_))
+
+    def test_void_scalar_empty_tuple(self):
+        s = np.zeros((), dtype='V4')
+        assert_equal(s[()].dtype, s.dtype)
+        assert_equal(s[()], s)
+        assert_equal(type(s[...]), np.ndarray)
+
+    def test_same_kind_index_casting(self):
+        # Indexes should be cast with same-kind and not safe, even if that
+        # is somewhat unsafe. So test various different code paths.
+        index = np.arange(5)
+        u_index = index.astype(np.uintp)
+        arr = np.arange(10)
+
+        assert_array_equal(arr[index], arr[u_index])
+        arr[u_index] = np.arange(5)
+        assert_array_equal(arr, np.arange(10))
+
+        arr = np.arange(10).reshape(5, 2)
+        assert_array_equal(arr[index], arr[u_index])
+
+        arr[u_index] = np.arange(5)[:,None]
+        assert_array_equal(arr, np.arange(5)[:,None].repeat(2, axis=1))
+
+        arr = np.arange(25).reshape(5, 5)
+        assert_array_equal(arr[u_index, u_index], arr[index, index])
+
+    def test_empty_fancy_index(self):
+        # Empty list index creates an empty array
+        # with the same dtype (but with weird shape)
+        a = np.array([1, 2, 3])
+        assert_equal(a[[]], [])
+        assert_equal(a[[]].dtype, a.dtype)
+
+        b = np.array([], dtype=np.intp)
+        assert_equal(a[[]], [])
+        assert_equal(a[[]].dtype, a.dtype)
+
+        b = np.array([])
+        assert_raises(IndexError, a.__getitem__, b)
+
+    def test_ellipsis_index(self):
+        a = np.array([[1, 2, 3],
+                      [4, 5, 6],
+                      [7, 8, 9]])
+        assert_(a[...] is not a)
+        assert_equal(a[...], a)
+        # `a[...]` was `a` in numpy <1.9.
+        assert_(a[...].base is a)
+
+        # Slicing with ellipsis can skip an
+        # arbitrary number of dimensions
+        assert_equal(a[0, ...], a[0])
+        assert_equal(a[0, ...], a[0,:])
+        assert_equal(a[..., 0], a[:, 0])
+
+        # Slicing with ellipsis always results
+        # in an array, not a scalar
+        assert_equal(a[0, ..., 1], np.array(2))
+
+        # Assignment with `(Ellipsis,)` on 0-d arrays
+        b = np.array(1)
+        b[(Ellipsis,)] = 2
+        assert_equal(b, 2)
+
+    def test_single_int_index(self):
+        # Single integer index selects one row
+        a = np.array([[1, 2, 3],
+                      [4, 5, 6],
+                      [7, 8, 9]])
+
+        assert_equal(a[0], [1, 2, 3])
+        assert_equal(a[-1], [7, 8, 9])
+
+        # Index out of bounds produces IndexError
+        assert_raises(IndexError, a.__getitem__, 1 << 30)
+        # Index overflow produces IndexError
+        assert_raises(IndexError, a.__getitem__, 1 << 64)
+
+    def test_single_bool_index(self):
+        # Single boolean index
+        a = np.array([[1, 2, 3],
+                      [4, 5, 6],
+                      [7, 8, 9]])
+
+        assert_equal(a[np.array(True)], a[None])
+        assert_equal(a[np.array(False)], a[None][0:0])
+
+    def test_boolean_shape_mismatch(self):
+        arr = np.ones((5, 4, 3))
+
+        index = np.array([True])
+        assert_raises(IndexError, arr.__getitem__, index)
+
+        index = np.array([False] * 6)
+        assert_raises(IndexError, arr.__getitem__, index)
+
+        index = np.zeros((4, 4), dtype=bool)
+        assert_raises(IndexError, arr.__getitem__, index)
+
+        assert_raises(IndexError, arr.__getitem__, (slice(None), index))
+
+    def test_boolean_indexing_onedim(self):
+        # Indexing a 2-dimensional array with
+        # boolean array of length one
+        a = np.array([[ 0.,  0.,  0.]])
+        b = np.array([ True], dtype=bool)
+        assert_equal(a[b], a)
+        # boolean assignment
+        a[b] = 1.
+        assert_equal(a, [[1., 1., 1.]])
+
+    def test_boolean_assignment_value_mismatch(self):
+        # A boolean assignment should fail when the shape of the values
+        # cannot be broadcast to the subscription. (see also gh-3458)
+        a = np.arange(4)
+
+        def f(a, v):
+            a[a > -1] = v
+
+        assert_raises(ValueError, f, a, [])
+        assert_raises(ValueError, f, a, [1, 2, 3])
+        assert_raises(ValueError, f, a[:1], [1, 2, 3])
+
+    def test_boolean_assignment_needs_api(self):
+        # See also gh-7666
+        # This caused a segfault on Python 2 due to the GIL not being
+        # held when the iterator does not need it, but the transfer function
+        # does
+        arr = np.zeros(1000)
+        indx = np.zeros(1000, dtype=bool)
+        indx[:100] = True
+        arr[indx] = np.ones(100, dtype=object)
+
+        expected = np.zeros(1000)
+        expected[:100] = 1
+        assert_array_equal(arr, expected)
+
+    def test_boolean_indexing_twodim(self):
+        # Indexing a 2-dimensional array with
+        # 2-dimensional boolean array
+        a = np.array([[1, 2, 3],
+                      [4, 5, 6],
+                      [7, 8, 9]])
+        b = np.array([[ True, False,  True],
+                      [False,  True, False],
+                      [ True, False,  True]])
+        assert_equal(a[b], [1, 3, 5, 7, 9])
+        assert_equal(a[b[1]], [[4, 5, 6]])
+        assert_equal(a[b[0]], a[b[2]])
+
+        # boolean assignment
+        a[b] = 0
+        assert_equal(a, [[0, 2, 0],
+                         [4, 0, 6],
+                         [0, 8, 0]])
+
+    def test_boolean_indexing_list(self):
+        # Regression test for #13715. It's a use-after-free bug which the
+        # test won't directly catch, but it will show up in valgrind.
+        a = np.array([1, 2, 3])
+        b = [True, False, True]
+        # Two variants of the test because the first takes a fast path
+        assert_equal(a[b], [1, 3])
+        assert_equal(a[None, b], [[1, 3]])
+
+    def test_reverse_strides_and_subspace_bufferinit(self):
+        # This tests that the strides are not reversed for simple and
+        # subspace fancy indexing.
+        a = np.ones(5)
+        b = np.zeros(5, dtype=np.intp)[::-1]
+        c = np.arange(5)[::-1]
+
+        a[b] = c
+        # If the strides are not reversed, the 0 in the arange comes last.
+        assert_equal(a[0], 0)
+
+        # This also tests that the subspace buffer is initialized:
+        a = np.ones((5, 2))
+        c = np.arange(10).reshape(5, 2)[::-1]
+        a[b, :] = c
+        assert_equal(a[0], [0, 1])
+
+    def test_reversed_strides_result_allocation(self):
+        # Test a bug when calculating the output strides for a result array
+        # when the subspace size was 1 (and test other cases as well)
+        a = np.arange(10)[:, None]
+        i = np.arange(10)[::-1]
+        assert_array_equal(a[i], a[i.copy('C')])
+
+        a = np.arange(20).reshape(-1, 2)
+
+    def test_uncontiguous_subspace_assignment(self):
+        # During development there was a bug activating a skip logic
+        # based on ndim instead of size.
+        a = np.full((3, 4, 2), -1)
+        b = np.full((3, 4, 2), -1)
+
+        a[[0, 1]] = np.arange(2 * 4 * 2).reshape(2, 4, 2).T
+        b[[0, 1]] = np.arange(2 * 4 * 2).reshape(2, 4, 2).T.copy()
+
+        assert_equal(a, b)
+
+    def test_too_many_fancy_indices_special_case(self):
+        # Just documents behaviour, this is a small limitation.
+        a = np.ones((1,) * 32)  # 32 is NPY_MAXDIMS
+        assert_raises(IndexError, a.__getitem__, (np.array([0]),) * 32)
+
+    def test_scalar_array_bool(self):
+        # NumPy bools can be used as boolean index (python ones as of yet not)
+        a = np.array(1)
+        assert_equal(a[np.bool_(True)], a[np.array(True)])
+        assert_equal(a[np.bool_(False)], a[np.array(False)])
+
+        # After deprecating bools as integers:
+        #a = np.array([0,1,2])
+        #assert_equal(a[True, :], a[None, :])
+        #assert_equal(a[:, True], a[:, None])
+        #
+        #assert_(not np.may_share_memory(a, a[True, :]))
+
+    def test_everything_returns_views(self):
+        # Before `...` would return a itself.
+        a = np.arange(5)
+
+        assert_(a is not a[()])
+        assert_(a is not a[...])
+        assert_(a is not a[:])
+
+    def test_broaderrors_indexing(self):
+        a = np.zeros((5, 5))
+        assert_raises(IndexError, a.__getitem__, ([0, 1], [0, 1, 2]))
+        assert_raises(IndexError, a.__setitem__, ([0, 1], [0, 1, 2]), 0)
+
+    def test_trivial_fancy_out_of_bounds(self):
+        a = np.zeros(5)
+        ind = np.ones(20, dtype=np.intp)
+        ind[-1] = 10
+        assert_raises(IndexError, a.__getitem__, ind)
+        assert_raises(IndexError, a.__setitem__, ind, 0)
+        ind = np.ones(20, dtype=np.intp)
+        ind[0] = 11
+        assert_raises(IndexError, a.__getitem__, ind)
+        assert_raises(IndexError, a.__setitem__, ind, 0)
+
+    def test_trivial_fancy_not_possible(self):
+        # Test that the fast path for trivial assignment is not incorrectly
+        # used when the index is not contiguous or 1D, see also gh-11467.
+        a = np.arange(6)
+        idx = np.arange(6, dtype=np.intp).reshape(2, 1, 3)[:, :, 0]
+        assert_array_equal(a[idx], idx)
+
+        # this case must not go into the fast path, note that idx is
+        # a non-contiuguous none 1D array here.
+        a[idx] = -1
+        res = np.arange(6)
+        res[0] = -1
+        res[3] = -1
+        assert_array_equal(a, res)
+
+    def test_nonbaseclass_values(self):
+        class SubClass(np.ndarray):
+            def __array_finalize__(self, old):
+                # Have array finalize do funny things
+                self.fill(99)
+
+        a = np.zeros((5, 5))
+        s = a.copy().view(type=SubClass)
+        s.fill(1)
+
+        a[[0, 1, 2, 3, 4], :] = s
+        assert_((a == 1).all())
+
+        # Subspace is last, so transposing might want to finalize
+        a[:, [0, 1, 2, 3, 4]] = s
+        assert_((a == 1).all())
+
+        a.fill(0)
+        a[...] = s
+        assert_((a == 1).all())
+
+    def test_array_like_values(self):
+        # Similar to the above test, but use a memoryview instead
+        a = np.zeros((5, 5))
+        s = np.arange(25, dtype=np.float64).reshape(5, 5)
+
+        a[[0, 1, 2, 3, 4], :] = memoryview(s)
+        assert_array_equal(a, s)
+
+        a[:, [0, 1, 2, 3, 4]] = memoryview(s)
+        assert_array_equal(a, s)
+
+        a[...] = memoryview(s)
+        assert_array_equal(a, s)
+
+    def test_subclass_writeable(self):
+        d = np.rec.array([('NGC1001', 11), ('NGC1002', 1.), ('NGC1003', 1.)],
+                         dtype=[('target', 'S20'), ('V_mag', '>f4')])
+        ind = np.array([False,  True,  True], dtype=bool)
+        assert_(d[ind].flags.writeable)
+        ind = np.array([0, 1])
+        assert_(d[ind].flags.writeable)
+        assert_(d[...].flags.writeable)
+        assert_(d[0].flags.writeable)
+
+    def test_memory_order(self):
+        # This is not necessary to preserve. Memory layouts for
+        # more complex indices are not as simple.
+        a = np.arange(10)
+        b = np.arange(10).reshape(5,2).T
+        assert_(a[b].flags.f_contiguous)
+
+        # Takes a different implementation branch:
+        a = a.reshape(-1, 1)
+        assert_(a[b, 0].flags.f_contiguous)
+
+    def test_scalar_return_type(self):
+        # Full scalar indices should return scalars and object
+        # arrays should not call PyArray_Return on their items
+        class Zero:
+            # The most basic valid indexing
+            def __index__(self):
+                return 0
+
+        z = Zero()
+
+        class ArrayLike:
+            # Simple array, should behave like the array
+            def __array__(self):
+                return np.array(0)
+
+        a = np.zeros(())
+        assert_(isinstance(a[()], np.float_))
+        a = np.zeros(1)
+        assert_(isinstance(a[z], np.float_))
+        a = np.zeros((1, 1))
+        assert_(isinstance(a[z, np.array(0)], np.float_))
+        assert_(isinstance(a[z, ArrayLike()], np.float_))
+
+        # And object arrays do not call it too often:
+        b = np.array(0)
+        a = np.array(0, dtype=object)
+        a[()] = b
+        assert_(isinstance(a[()], np.ndarray))
+        a = np.array([b, None])
+        assert_(isinstance(a[z], np.ndarray))
+        a = np.array([[b, None]])
+        assert_(isinstance(a[z, np.array(0)], np.ndarray))
+        assert_(isinstance(a[z, ArrayLike()], np.ndarray))
+
+    def test_small_regressions(self):
+        # Reference count of intp for index checks
+        a = np.array([0])
+        if HAS_REFCOUNT:
+            refcount = sys.getrefcount(np.dtype(np.intp))
+        # item setting always checks indices in separate function:
+        a[np.array([0], dtype=np.intp)] = 1
+        a[np.array([0], dtype=np.uint8)] = 1
+        assert_raises(IndexError, a.__setitem__,
+                      np.array([1], dtype=np.intp), 1)
+        assert_raises(IndexError, a.__setitem__,
+                      np.array([1], dtype=np.uint8), 1)
+
+        if HAS_REFCOUNT:
+            assert_equal(sys.getrefcount(np.dtype(np.intp)), refcount)
+
+    def test_unaligned(self):
+        v = (np.zeros(64, dtype=np.int8) + ord('a'))[1:-7]
+        d = v.view(np.dtype("S8"))
+        # unaligned source
+        x = (np.zeros(16, dtype=np.int8) + ord('a'))[1:-7]
+        x = x.view(np.dtype("S8"))
+        x[...] = np.array("b" * 8, dtype="S")
+        b = np.arange(d.size)
+        #trivial
+        assert_equal(d[b], d)
+        d[b] = x
+        # nontrivial
+        # unaligned index array
+        b = np.zeros(d.size + 1).view(np.int8)[1:-(np.intp(0).itemsize - 1)]
+        b = b.view(np.intp)[:d.size]
+        b[...] = np.arange(d.size)
+        assert_equal(d[b.astype(np.int16)], d)
+        d[b.astype(np.int16)] = x
+        # boolean
+        d[b % 2 == 0]
+        d[b % 2 == 0] = x[::2]
+
+    def test_tuple_subclass(self):
+        arr = np.ones((5, 5))
+
+        # A tuple subclass should also be an nd-index
+        class TupleSubclass(tuple):
+            pass
+        index = ([1], [1])
+        index = TupleSubclass(index)
+        assert_(arr[index].shape == (1,))
+        # Unlike the non nd-index:
+        assert_(arr[index,].shape != (1,))
+
+    def test_broken_sequence_not_nd_index(self):
+        # See gh-5063:
+        # If we have an object which claims to be a sequence, but fails
+        # on item getting, this should not be converted to an nd-index (tuple)
+        # If this object happens to be a valid index otherwise, it should work
+        # This object here is very dubious and probably bad though:
+        class SequenceLike:
+            def __index__(self):
+                return 0
+
+            def __len__(self):
+                return 1
+
+            def __getitem__(self, item):
+                raise IndexError('Not possible')
+
+        arr = np.arange(10)
+        assert_array_equal(arr[SequenceLike()], arr[SequenceLike(),])
+
+        # also test that field indexing does not segfault
+        # for a similar reason, by indexing a structured array
+        arr = np.zeros((1,), dtype=[('f1', 'i8'), ('f2', 'i8')])
+        assert_array_equal(arr[SequenceLike()], arr[SequenceLike(),])
+
+    def test_indexing_array_weird_strides(self):
+        # See also gh-6221
+        # the shapes used here come from the issue and create the correct
+        # size for the iterator buffering size.
+        x = np.ones(10)
+        x2 = np.ones((10, 2))
+        ind = np.arange(10)[:, None, None, None]
+        ind = np.broadcast_to(ind, (10, 55, 4, 4))
+
+        # single advanced index case
+        assert_array_equal(x[ind], x[ind.copy()])
+        # higher dimensional advanced index
+        zind = np.zeros(4, dtype=np.intp)
+        assert_array_equal(x2[ind, zind], x2[ind.copy(), zind])
+
+    def test_indexing_array_negative_strides(self):
+        # From gh-8264,
+        # core dumps if negative strides are used in iteration
+        arro = np.zeros((4, 4))
+        arr = arro[::-1, ::-1]
+
+        slices = (slice(None), [0, 1, 2, 3])
+        arr[slices] = 10
+        assert_array_equal(arr, 10.)
+
+    def test_character_assignment(self):
+        # This is an example a function going through CopyObject which
+        # used to have an untested special path for scalars
+        # (the character special dtype case, should be deprecated probably)
+        arr = np.zeros((1, 5), dtype="c")
+        arr[0] = np.str_("asdfg")  # must assign as a sequence
+        assert_array_equal(arr[0], np.array("asdfg", dtype="c"))
+        assert arr[0, 1] == b"s"  # make sure not all were set to "a" for both
+
+    @pytest.mark.parametrize("index",
+            [True, False, np.array([0])])
+    @pytest.mark.parametrize("num", [32, 40])
+    @pytest.mark.parametrize("original_ndim", [1, 32])
+    def test_too_many_advanced_indices(self, index, num, original_ndim):
+        # These are limitations based on the number of arguments we can process.
+        # For `num=32` (and all boolean cases), the result is actually define;
+        # but the use of NpyIter (NPY_MAXARGS) limits it for technical reasons.
+        arr = np.ones((1,) * original_ndim)
+        with pytest.raises(IndexError):
+            arr[(index,) * num]
+        with pytest.raises(IndexError):
+            arr[(index,) * num] = 1.
+
+    @pytest.mark.skipif(IS_WASM, reason="no threading")
+    def test_structured_advanced_indexing(self):
+        # Test that copyswap(n) used by integer array indexing is threadsafe
+        # for structured datatypes, see gh-15387. This test can behave randomly.
+        from concurrent.futures import ThreadPoolExecutor
+
+        # Create a deeply nested dtype to make a failure more likely:
+        dt = np.dtype([("", "f8")])
+        dt = np.dtype([("", dt)] * 2)
+        dt = np.dtype([("", dt)] * 2)
+        # The array should be large enough to likely run into threading issues
+        arr = np.random.uniform(size=(6000, 8)).view(dt)[:, 0]
+
+        rng = np.random.default_rng()
+        def func(arr):
+            indx = rng.integers(0, len(arr), size=6000, dtype=np.intp)
+            arr[indx]
+
+        tpe = ThreadPoolExecutor(max_workers=8)
+        futures = [tpe.submit(func, arr) for _ in range(10)]
+        for f in futures:
+            f.result()
+
+        assert arr.dtype is dt
+
+    def test_nontuple_ndindex(self):
+        a = np.arange(25).reshape((5, 5))
+        assert_equal(a[[0, 1]], np.array([a[0], a[1]]))
+        assert_equal(a[[0, 1], [0, 1]], np.array([0, 6]))
+        assert_raises(IndexError, a.__getitem__, [slice(None)])
+
+
+class TestFieldIndexing:
+    def test_scalar_return_type(self):
+        # Field access on an array should return an array, even if it
+        # is 0-d.
+        a = np.zeros((), [('a','f8')])
+        assert_(isinstance(a['a'], np.ndarray))
+        assert_(isinstance(a[['a']], np.ndarray))
+
+
+class TestBroadcastedAssignments:
+    def assign(self, a, ind, val):
+        a[ind] = val
+        return a
+
+    def test_prepending_ones(self):
+        a = np.zeros((3, 2))
+
+        a[...] = np.ones((1, 3, 2))
+        # Fancy with subspace with and without transpose
+        a[[0, 1, 2], :] = np.ones((1, 3, 2))
+        a[:, [0, 1]] = np.ones((1, 3, 2))
+        # Fancy without subspace (with broadcasting)
+        a[[[0], [1], [2]], [0, 1]] = np.ones((1, 3, 2))
+
+    def test_prepend_not_one(self):
+        assign = self.assign
+        s_ = np.s_
+        a = np.zeros(5)
+
+        # Too large and not only ones.
+        assert_raises(ValueError, assign, a, s_[...],  np.ones((2, 1)))
+        assert_raises(ValueError, assign, a, s_[[1, 2, 3],], np.ones((2, 1)))
+        assert_raises(ValueError, assign, a, s_[[[1], [2]],], np.ones((2,2,1)))
+
+    def test_simple_broadcasting_errors(self):
+        assign = self.assign
+        s_ = np.s_
+        a = np.zeros((5, 1))
+
+        assert_raises(ValueError, assign, a, s_[...], np.zeros((5, 2)))
+        assert_raises(ValueError, assign, a, s_[...], np.zeros((5, 0)))
+        assert_raises(ValueError, assign, a, s_[:, [0]], np.zeros((5, 2)))
+        assert_raises(ValueError, assign, a, s_[:, [0]], np.zeros((5, 0)))
+        assert_raises(ValueError, assign, a, s_[[0], :], np.zeros((2, 1)))
+
+    @pytest.mark.parametrize("index", [
+            (..., [1, 2], slice(None)),
+            ([0, 1], ..., 0),
+            (..., [1, 2], [1, 2])])
+    def test_broadcast_error_reports_correct_shape(self, index):
+        values = np.zeros((100, 100))  # will never broadcast below  
+
+        arr = np.zeros((3, 4, 5, 6, 7))
+        # We currently report without any spaces (could be changed)
+        shape_str = str(arr[index].shape).replace(" ", "")
+        
+        with pytest.raises(ValueError) as e:
+            arr[index] = values
+
+        assert str(e.value).endswith(shape_str)
+
+    def test_index_is_larger(self):
+        # Simple case of fancy index broadcasting of the index.
+        a = np.zeros((5, 5))
+        a[[[0], [1], [2]], [0, 1, 2]] = [2, 3, 4]
+
+        assert_((a[:3, :3] == [2, 3, 4]).all())
+
+    def test_broadcast_subspace(self):
+        a = np.zeros((100, 100))
+        v = np.arange(100)[:,None]
+        b = np.arange(100)[::-1]
+        a[b] = v
+        assert_((a[::-1] == v).all())
+
+
+class TestSubclasses:
+    def test_basic(self):
+        # Test that indexing in various ways produces SubClass instances,
+        # and that the base is set up correctly: the original subclass
+        # instance for views, and a new ndarray for advanced/boolean indexing
+        # where a copy was made (latter a regression test for gh-11983).
+        class SubClass(np.ndarray):
+            pass
+
+        a = np.arange(5)
+        s = a.view(SubClass)
+        s_slice = s[:3]
+        assert_(type(s_slice) is SubClass)
+        assert_(s_slice.base is s)
+        assert_array_equal(s_slice, a[:3])
+
+        s_fancy = s[[0, 1, 2]]
+        assert_(type(s_fancy) is SubClass)
+        assert_(s_fancy.base is not s)
+        assert_(type(s_fancy.base) is np.ndarray)
+        assert_array_equal(s_fancy, a[[0, 1, 2]])
+        assert_array_equal(s_fancy.base, a[[0, 1, 2]])
+
+        s_bool = s[s > 0]
+        assert_(type(s_bool) is SubClass)
+        assert_(s_bool.base is not s)
+        assert_(type(s_bool.base) is np.ndarray)
+        assert_array_equal(s_bool, a[a > 0])
+        assert_array_equal(s_bool.base, a[a > 0])
+
+    def test_fancy_on_read_only(self):
+        # Test that fancy indexing on read-only SubClass does not make a
+        # read-only copy (gh-14132)
+        class SubClass(np.ndarray):
+            pass
+
+        a = np.arange(5)
+        s = a.view(SubClass)
+        s.flags.writeable = False
+        s_fancy = s[[0, 1, 2]]
+        assert_(s_fancy.flags.writeable)
+
+
+    def test_finalize_gets_full_info(self):
+        # Array finalize should be called on the filled array.
+        class SubClass(np.ndarray):
+            def __array_finalize__(self, old):
+                self.finalize_status = np.array(self)
+                self.old = old
+
+        s = np.arange(10).view(SubClass)
+        new_s = s[:3]
+        assert_array_equal(new_s.finalize_status, new_s)
+        assert_array_equal(new_s.old, s)
+
+        new_s = s[[0,1,2,3]]
+        assert_array_equal(new_s.finalize_status, new_s)
+        assert_array_equal(new_s.old, s)
+
+        new_s = s[s > 0]
+        assert_array_equal(new_s.finalize_status, new_s)
+        assert_array_equal(new_s.old, s)
+
+
+class TestFancyIndexingCast:
+    def test_boolean_index_cast_assign(self):
+        # Setup the boolean index and float arrays.
+        shape = (8, 63)
+        bool_index = np.zeros(shape).astype(bool)
+        bool_index[0, 1] = True
+        zero_array = np.zeros(shape)
+
+        # Assigning float is fine.
+        zero_array[bool_index] = np.array([1])
+        assert_equal(zero_array[0, 1], 1)
+
+        # Fancy indexing works, although we get a cast warning.
+        assert_warns(np.ComplexWarning,
+                     zero_array.__setitem__, ([0], [1]), np.array([2 + 1j]))
+        assert_equal(zero_array[0, 1], 2)  # No complex part
+
+        # Cast complex to float, throwing away the imaginary portion.
+        assert_warns(np.ComplexWarning,
+                     zero_array.__setitem__, bool_index, np.array([1j]))
+        assert_equal(zero_array[0, 1], 0)
+
+class TestFancyIndexingEquivalence:
+    def test_object_assign(self):
+        # Check that the field and object special case using copyto is active.
+        # The right hand side cannot be converted to an array here.
+        a = np.arange(5, dtype=object)
+        b = a.copy()
+        a[:3] = [1, (1,2), 3]
+        b[[0, 1, 2]] = [1, (1,2), 3]
+        assert_array_equal(a, b)
+
+        # test same for subspace fancy indexing
+        b = np.arange(5, dtype=object)[None, :]
+        b[[0], :3] = [[1, (1,2), 3]]
+        assert_array_equal(a, b[0])
+
+        # Check that swapping of axes works.
+        # There was a bug that made the later assignment throw a ValueError
+        # do to an incorrectly transposed temporary right hand side (gh-5714)
+        b = b.T
+        b[:3, [0]] = [[1], [(1,2)], [3]]
+        assert_array_equal(a, b[:, 0])
+
+        # Another test for the memory order of the subspace
+        arr = np.ones((3, 4, 5), dtype=object)
+        # Equivalent slicing assignment for comparison
+        cmp_arr = arr.copy()
+        cmp_arr[:1, ...] = [[[1], [2], [3], [4]]]
+        arr[[0], ...] = [[[1], [2], [3], [4]]]
+        assert_array_equal(arr, cmp_arr)
+        arr = arr.copy('F')
+        arr[[0], ...] = [[[1], [2], [3], [4]]]
+        assert_array_equal(arr, cmp_arr)
+
+    def test_cast_equivalence(self):
+        # Yes, normal slicing uses unsafe casting.
+        a = np.arange(5)
+        b = a.copy()
+
+        a[:3] = np.array(['2', '-3', '-1'])
+        b[[0, 2, 1]] = np.array(['2', '-1', '-3'])
+        assert_array_equal(a, b)
+
+        # test the same for subspace fancy indexing
+        b = np.arange(5)[None, :]
+        b[[0], :3] = np.array([['2', '-3', '-1']])
+        assert_array_equal(a, b[0])
+
+
+class TestMultiIndexingAutomated:
+    """
+    These tests use code to mimic the C-Code indexing for selection.
+
+    NOTE:
+
+        * This still lacks tests for complex item setting.
+        * If you change behavior of indexing, you might want to modify
+          these tests to try more combinations.
+        * Behavior was written to match numpy version 1.8. (though a
+          first version matched 1.7.)
+        * Only tuple indices are supported by the mimicking code.
+          (and tested as of writing this)
+        * Error types should match most of the time as long as there
+          is only one error. For multiple errors, what gets raised
+          will usually not be the same one. They are *not* tested.
+
+    Update 2016-11-30: It is probably not worth maintaining this test
+    indefinitely and it can be dropped if maintenance becomes a burden.
+
+    """
+
+    def setup_method(self):
+        self.a = np.arange(np.prod([3, 1, 5, 6])).reshape(3, 1, 5, 6)
+        self.b = np.empty((3, 0, 5, 6))
+        self.complex_indices = ['skip', Ellipsis,
+            0,
+            # Boolean indices, up to 3-d for some special cases of eating up
+            # dimensions, also need to test all False
+            np.array([True, False, False]),
+            np.array([[True, False], [False, True]]),
+            np.array([[[False, False], [False, False]]]),
+            # Some slices:
+            slice(-5, 5, 2),
+            slice(1, 1, 100),
+            slice(4, -1, -2),
+            slice(None, None, -3),
+            # Some Fancy indexes:
+            np.empty((0, 1, 1), dtype=np.intp),  # empty and can be broadcast
+            np.array([0, 1, -2]),
+            np.array([[2], [0], [1]]),
+            np.array([[0, -1], [0, 1]], dtype=np.dtype('intp').newbyteorder()),
+            np.array([2, -1], dtype=np.int8),
+            np.zeros([1]*31, dtype=int),  # trigger too large array.
+            np.array([0., 1.])]  # invalid datatype
+        # Some simpler indices that still cover a bit more
+        self.simple_indices = [Ellipsis, None, -1, [1], np.array([True]),
+                               'skip']
+        # Very simple ones to fill the rest:
+        self.fill_indices = [slice(None, None), 0]
+
+    def _get_multi_index(self, arr, indices):
+        """Mimic multi dimensional indexing.
+
+        Parameters
+        ----------
+        arr : ndarray
+            Array to be indexed.
+        indices : tuple of index objects
+
+        Returns
+        -------
+        out : ndarray
+            An array equivalent to the indexing operation (but always a copy).
+            `arr[indices]` should be identical.
+        no_copy : bool
+            Whether the indexing operation requires a copy. If this is `True`,
+            `np.may_share_memory(arr, arr[indices])` should be `True` (with
+            some exceptions for scalars and possibly 0-d arrays).
+
+        Notes
+        -----
+        While the function may mostly match the errors of normal indexing this
+        is generally not the case.
+        """
+        in_indices = list(indices)
+        indices = []
+        # if False, this is a fancy or boolean index
+        no_copy = True
+        # number of fancy/scalar indexes that are not consecutive
+        num_fancy = 0
+        # number of dimensions indexed by a "fancy" index
+        fancy_dim = 0
+        # NOTE: This is a funny twist (and probably OK to change).
+        # The boolean array has illegal indexes, but this is
+        # allowed if the broadcast fancy-indices are 0-sized.
+        # This variable is to catch that case.
+        error_unless_broadcast_to_empty = False
+
+        # We need to handle Ellipsis and make arrays from indices, also
+        # check if this is fancy indexing (set no_copy).
+        ndim = 0
+        ellipsis_pos = None  # define here mostly to replace all but first.
+        for i, indx in enumerate(in_indices):
+            if indx is None:
+                continue
+            if isinstance(indx, np.ndarray) and indx.dtype == bool:
+                no_copy = False
+                if indx.ndim == 0:
+                    raise IndexError
+                # boolean indices can have higher dimensions
+                ndim += indx.ndim
+                fancy_dim += indx.ndim
+                continue
+            if indx is Ellipsis:
+                if ellipsis_pos is None:
+                    ellipsis_pos = i
+                    continue  # do not increment ndim counter
+                raise IndexError
+            if isinstance(indx, slice):
+                ndim += 1
+                continue
+            if not isinstance(indx, np.ndarray):
+                # This could be open for changes in numpy.
+                # numpy should maybe raise an error if casting to intp
+                # is not safe. It rejects np.array([1., 2.]) but not
+                # [1., 2.] as index (same for ie. np.take).
+                # (Note the importance of empty lists if changing this here)
+                try:
+                    indx = np.array(indx, dtype=np.intp)
+                except ValueError:
+                    raise IndexError
+                in_indices[i] = indx
+            elif indx.dtype.kind != 'b' and indx.dtype.kind != 'i':
+                raise IndexError('arrays used as indices must be of '
+                                 'integer (or boolean) type')
+            if indx.ndim != 0:
+                no_copy = False
+            ndim += 1
+            fancy_dim += 1
+
+        if arr.ndim - ndim < 0:
+            # we can't take more dimensions then we have, not even for 0-d
+            # arrays.  since a[()] makes sense, but not a[(),]. We will
+            # raise an error later on, unless a broadcasting error occurs
+            # first.
+            raise IndexError
+
+        if ndim == 0 and None not in in_indices:
+            # Well we have no indexes or one Ellipsis. This is legal.
+            return arr.copy(), no_copy
+
+        if ellipsis_pos is not None:
+            in_indices[ellipsis_pos:ellipsis_pos+1] = ([slice(None, None)] *
+                                                       (arr.ndim - ndim))
+
+        for ax, indx in enumerate(in_indices):
+            if isinstance(indx, slice):
+                # convert to an index array
+                indx = np.arange(*indx.indices(arr.shape[ax]))
+                indices.append(['s', indx])
+                continue
+            elif indx is None:
+                # this is like taking a slice with one element from a new axis:
+                indices.append(['n', np.array([0], dtype=np.intp)])
+                arr = arr.reshape((arr.shape[:ax] + (1,) + arr.shape[ax:]))
+                continue
+            if isinstance(indx, np.ndarray) and indx.dtype == bool:
+                if indx.shape != arr.shape[ax:ax+indx.ndim]:
+                    raise IndexError
+
+                try:
+                    flat_indx = np.ravel_multi_index(np.nonzero(indx),
+                                    arr.shape[ax:ax+indx.ndim], mode='raise')
+                except Exception:
+                    error_unless_broadcast_to_empty = True
+                    # fill with 0s instead, and raise error later
+                    flat_indx = np.array([0]*indx.sum(), dtype=np.intp)
+                # concatenate axis into a single one:
+                if indx.ndim != 0:
+                    arr = arr.reshape((arr.shape[:ax]
+                                  + (np.prod(arr.shape[ax:ax+indx.ndim]),)
+                                  + arr.shape[ax+indx.ndim:]))
+                    indx = flat_indx
+                else:
+                    # This could be changed, a 0-d boolean index can
+                    # make sense (even outside the 0-d indexed array case)
+                    # Note that originally this is could be interpreted as
+                    # integer in the full integer special case.
+                    raise IndexError
+            else:
+                # If the index is a singleton, the bounds check is done
+                # before the broadcasting. This used to be different in <1.9
+                if indx.ndim == 0:
+                    if indx >= arr.shape[ax] or indx < -arr.shape[ax]:
+                        raise IndexError
+            if indx.ndim == 0:
+                # The index is a scalar. This used to be two fold, but if
+                # fancy indexing was active, the check was done later,
+                # possibly after broadcasting it away (1.7. or earlier).
+                # Now it is always done.
+                if indx >= arr.shape[ax] or indx < - arr.shape[ax]:
+                    raise IndexError
+            if (len(indices) > 0 and
+                    indices[-1][0] == 'f' and
+                    ax != ellipsis_pos):
+                # NOTE: There could still have been a 0-sized Ellipsis
+                # between them. Checked that with ellipsis_pos.
+                indices[-1].append(indx)
+            else:
+                # We have a fancy index that is not after an existing one.
+                # NOTE: A 0-d array triggers this as well, while one may
+                # expect it to not trigger it, since a scalar would not be
+                # considered fancy indexing.
+                num_fancy += 1
+                indices.append(['f', indx])
+
+        if num_fancy > 1 and not no_copy:
+            # We have to flush the fancy indexes left
+            new_indices = indices[:]
+            axes = list(range(arr.ndim))
+            fancy_axes = []
+            new_indices.insert(0, ['f'])
+            ni = 0
+            ai = 0
+            for indx in indices:
+                ni += 1
+                if indx[0] == 'f':
+                    new_indices[0].extend(indx[1:])
+                    del new_indices[ni]
+                    ni -= 1
+                    for ax in range(ai, ai + len(indx[1:])):
+                        fancy_axes.append(ax)
+                        axes.remove(ax)
+                ai += len(indx) - 1  # axis we are at
+            indices = new_indices
+            # and now we need to transpose arr:
+            arr = arr.transpose(*(fancy_axes + axes))
+
+        # We only have one 'f' index now and arr is transposed accordingly.
+        # Now handle newaxis by reshaping...
+        ax = 0
+        for indx in indices:
+            if indx[0] == 'f':
+                if len(indx) == 1:
+                    continue
+                # First of all, reshape arr to combine fancy axes into one:
+                orig_shape = arr.shape
+                orig_slice = orig_shape[ax:ax + len(indx[1:])]
+                arr = arr.reshape((arr.shape[:ax]
+                                    + (np.prod(orig_slice).astype(int),)
+                                    + arr.shape[ax + len(indx[1:]):]))
+
+                # Check if broadcasting works
+                res = np.broadcast(*indx[1:])
+                # unfortunately the indices might be out of bounds. So check
+                # that first, and use mode='wrap' then. However only if
+                # there are any indices...
+                if res.size != 0:
+                    if error_unless_broadcast_to_empty:
+                        raise IndexError
+                    for _indx, _size in zip(indx[1:], orig_slice):
+                        if _indx.size == 0:
+                            continue
+                        if np.any(_indx >= _size) or np.any(_indx < -_size):
+                                raise IndexError
+                if len(indx[1:]) == len(orig_slice):
+                    if np.prod(orig_slice) == 0:
+                        # Work around for a crash or IndexError with 'wrap'
+                        # in some 0-sized cases.
+                        try:
+                            mi = np.ravel_multi_index(indx[1:], orig_slice,
+                                                      mode='raise')
+                        except Exception:
+                            # This happens with 0-sized orig_slice (sometimes?)
+                            # here it is a ValueError, but indexing gives a:
+                            raise IndexError('invalid index into 0-sized')
+                    else:
+                        mi = np.ravel_multi_index(indx[1:], orig_slice,
+                                                  mode='wrap')
+                else:
+                    # Maybe never happens...
+                    raise ValueError
+                arr = arr.take(mi.ravel(), axis=ax)
+                try:
+                    arr = arr.reshape((arr.shape[:ax]
+                                        + mi.shape
+                                        + arr.shape[ax+1:]))
+                except ValueError:
+                    # too many dimensions, probably
+                    raise IndexError
+                ax += mi.ndim
+                continue
+
+            # If we are here, we have a 1D array for take:
+            arr = arr.take(indx[1], axis=ax)
+            ax += 1
+
+        return arr, no_copy
+
+    def _check_multi_index(self, arr, index):
+        """Check a multi index item getting and simple setting.
+
+        Parameters
+        ----------
+        arr : ndarray
+            Array to be indexed, must be a reshaped arange.
+        index : tuple of indexing objects
+            Index being tested.
+        """
+        # Test item getting
+        try:
+            mimic_get, no_copy = self._get_multi_index(arr, index)
+        except Exception as e:
+            if HAS_REFCOUNT:
+                prev_refcount = sys.getrefcount(arr)
+            assert_raises(type(e), arr.__getitem__, index)
+            assert_raises(type(e), arr.__setitem__, index, 0)
+            if HAS_REFCOUNT:
+                assert_equal(prev_refcount, sys.getrefcount(arr))
+            return
+
+        self._compare_index_result(arr, index, mimic_get, no_copy)
+
+    def _check_single_index(self, arr, index):
+        """Check a single index item getting and simple setting.
+
+        Parameters
+        ----------
+        arr : ndarray
+            Array to be indexed, must be an arange.
+        index : indexing object
+            Index being tested. Must be a single index and not a tuple
+            of indexing objects (see also `_check_multi_index`).
+        """
+        try:
+            mimic_get, no_copy = self._get_multi_index(arr, (index,))
+        except Exception as e:
+            if HAS_REFCOUNT:
+                prev_refcount = sys.getrefcount(arr)
+            assert_raises(type(e), arr.__getitem__, index)
+            assert_raises(type(e), arr.__setitem__, index, 0)
+            if HAS_REFCOUNT:
+                assert_equal(prev_refcount, sys.getrefcount(arr))
+            return
+
+        self._compare_index_result(arr, index, mimic_get, no_copy)
+
+    def _compare_index_result(self, arr, index, mimic_get, no_copy):
+        """Compare mimicked result to indexing result.
+        """
+        arr = arr.copy()
+        indexed_arr = arr[index]
+        assert_array_equal(indexed_arr, mimic_get)
+        # Check if we got a view, unless its a 0-sized or 0-d array.
+        # (then its not a view, and that does not matter)
+        if indexed_arr.size != 0 and indexed_arr.ndim != 0:
+            assert_(np.may_share_memory(indexed_arr, arr) == no_copy)
+            # Check reference count of the original array
+            if HAS_REFCOUNT:
+                if no_copy:
+                    # refcount increases by one:
+                    assert_equal(sys.getrefcount(arr), 3)
+                else:
+                    assert_equal(sys.getrefcount(arr), 2)
+
+        # Test non-broadcast setitem:
+        b = arr.copy()
+        b[index] = mimic_get + 1000
+        if b.size == 0:
+            return  # nothing to compare here...
+        if no_copy and indexed_arr.ndim != 0:
+            # change indexed_arr in-place to manipulate original:
+            indexed_arr += 1000
+            assert_array_equal(arr, b)
+            return
+        # Use the fact that the array is originally an arange:
+        arr.flat[indexed_arr.ravel()] += 1000
+        assert_array_equal(arr, b)
+
+    def test_boolean(self):
+        a = np.array(5)
+        assert_equal(a[np.array(True)], 5)
+        a[np.array(True)] = 1
+        assert_equal(a, 1)
+        # NOTE: This is different from normal broadcasting, as
+        # arr[boolean_array] works like in a multi index. Which means
+        # it is aligned to the left. This is probably correct for
+        # consistency with arr[boolean_array,] also no broadcasting
+        # is done at all
+        self._check_multi_index(
+            self.a, (np.zeros_like(self.a, dtype=bool),))
+        self._check_multi_index(
+            self.a, (np.zeros_like(self.a, dtype=bool)[..., 0],))
+        self._check_multi_index(
+            self.a, (np.zeros_like(self.a, dtype=bool)[None, ...],))
+
+    def test_multidim(self):
+        # Automatically test combinations with complex indexes on 2nd (or 1st)
+        # spot and the simple ones in one other spot.
+        with warnings.catch_warnings():
+            # This is so that np.array(True) is not accepted in a full integer
+            # index, when running the file separately.
+            warnings.filterwarnings('error', '', DeprecationWarning)
+            warnings.filterwarnings('error', '', np.VisibleDeprecationWarning)
+
+            def isskip(idx):
+                return isinstance(idx, str) and idx == "skip"
+
+            for simple_pos in [0, 2, 3]:
+                tocheck = [self.fill_indices, self.complex_indices,
+                           self.fill_indices, self.fill_indices]
+                tocheck[simple_pos] = self.simple_indices
+                for index in product(*tocheck):
+                    index = tuple(i for i in index if not isskip(i))
+                    self._check_multi_index(self.a, index)
+                    self._check_multi_index(self.b, index)
+
+        # Check very simple item getting:
+        self._check_multi_index(self.a, (0, 0, 0, 0))
+        self._check_multi_index(self.b, (0, 0, 0, 0))
+        # Also check (simple cases of) too many indices:
+        assert_raises(IndexError, self.a.__getitem__, (0, 0, 0, 0, 0))
+        assert_raises(IndexError, self.a.__setitem__, (0, 0, 0, 0, 0), 0)
+        assert_raises(IndexError, self.a.__getitem__, (0, 0, [1], 0, 0))
+        assert_raises(IndexError, self.a.__setitem__, (0, 0, [1], 0, 0), 0)
+
+    def test_1d(self):
+        a = np.arange(10)
+        for index in self.complex_indices:
+            self._check_single_index(a, index)
+
+class TestFloatNonIntegerArgument:
+    """
+    These test that ``TypeError`` is raised when you try to use
+    non-integers as arguments to for indexing and slicing e.g. ``a[0.0:5]``
+    and ``a[0.5]``, or other functions like ``array.reshape(1., -1)``.
+
+    """
+    def test_valid_indexing(self):
+        # These should raise no errors.
+        a = np.array([[[5]]])
+
+        a[np.array([0])]
+        a[[0, 0]]
+        a[:, [0, 0]]
+        a[:, 0,:]
+        a[:,:,:]
+
+    def test_valid_slicing(self):
+        # These should raise no errors.
+        a = np.array([[[5]]])
+
+        a[::]
+        a[0:]
+        a[:2]
+        a[0:2]
+        a[::2]
+        a[1::2]
+        a[:2:2]
+        a[1:2:2]
+
+    def test_non_integer_argument_errors(self):
+        a = np.array([[5]])
+
+        assert_raises(TypeError, np.reshape, a, (1., 1., -1))
+        assert_raises(TypeError, np.reshape, a, (np.array(1.), -1))
+        assert_raises(TypeError, np.take, a, [0], 1.)
+        assert_raises(TypeError, np.take, a, [0], np.float64(1.))
+
+    def test_non_integer_sequence_multiplication(self):
+        # NumPy scalar sequence multiply should not work with non-integers
+        def mult(a, b):
+            return a * b
+
+        assert_raises(TypeError, mult, [1], np.float_(3))
+        # following should be OK
+        mult([1], np.int_(3))
+
+    def test_reduce_axis_float_index(self):
+        d = np.zeros((3,3,3))
+        assert_raises(TypeError, np.min, d, 0.5)
+        assert_raises(TypeError, np.min, d, (0.5, 1))
+        assert_raises(TypeError, np.min, d, (1, 2.2))
+        assert_raises(TypeError, np.min, d, (.2, 1.2))
+
+
+class TestBooleanIndexing:
+    # Using a boolean as integer argument/indexing is an error.
+    def test_bool_as_int_argument_errors(self):
+        a = np.array([[[1]]])
+
+        assert_raises(TypeError, np.reshape, a, (True, -1))
+        assert_raises(TypeError, np.reshape, a, (np.bool_(True), -1))
+        # Note that operator.index(np.array(True)) does not work, a boolean
+        # array is thus also deprecated, but not with the same message:
+        assert_raises(TypeError, operator.index, np.array(True))
+        assert_warns(DeprecationWarning, operator.index, np.True_)
+        assert_raises(TypeError, np.take, args=(a, [0], False))
+
+    def test_boolean_indexing_weirdness(self):
+        # Weird boolean indexing things
+        a = np.ones((2, 3, 4))
+        assert a[False, True, ...].shape == (0, 2, 3, 4)
+        assert a[True, [0, 1], True, True, [1], [[2]]].shape == (1, 2)
+        assert_raises(IndexError, lambda: a[False, [0, 1], ...])
+
+    def test_boolean_indexing_fast_path(self):
+        # These used to either give the wrong error, or incorrectly give no
+        # error.
+        a = np.ones((3, 3))
+
+        # This used to incorrectly work (and give an array of shape (0,))
+        idx1 = np.array([[False]*9])
+        assert_raises_regex(IndexError,
+            "boolean index did not match indexed array along dimension 0; "
+            "dimension is 3 but corresponding boolean dimension is 1",
+            lambda: a[idx1])
+
+        # This used to incorrectly give a ValueError: operands could not be broadcast together
+        idx2 = np.array([[False]*8 + [True]])
+        assert_raises_regex(IndexError,
+            "boolean index did not match indexed array along dimension 0; "
+            "dimension is 3 but corresponding boolean dimension is 1",
+            lambda: a[idx2])
+
+        # This is the same as it used to be. The above two should work like this.
+        idx3 = np.array([[False]*10])
+        assert_raises_regex(IndexError,
+            "boolean index did not match indexed array along dimension 0; "
+            "dimension is 3 but corresponding boolean dimension is 1",
+            lambda: a[idx3])
+
+        # This used to give ValueError: non-broadcastable operand
+        a = np.ones((1, 1, 2))
+        idx = np.array([[[True], [False]]])
+        assert_raises_regex(IndexError,
+            "boolean index did not match indexed array along dimension 1; "
+            "dimension is 1 but corresponding boolean dimension is 2",
+            lambda: a[idx])
+
+
+class TestArrayToIndexDeprecation:
+    """Creating an index from array not 0-D is an error.
+
+    """
+    def test_array_to_index_error(self):
+        # so no exception is expected. The raising is effectively tested above.
+        a = np.array([[[1]]])
+
+        assert_raises(TypeError, operator.index, np.array([1]))
+        assert_raises(TypeError, np.reshape, a, (a, -1))
+        assert_raises(TypeError, np.take, a, [0], a)
+
+
+class TestNonIntegerArrayLike:
+    """Tests that array_likes only valid if can safely cast to integer.
+
+    For instance, lists give IndexError when they cannot be safely cast to
+    an integer.
+
+    """
+    def test_basic(self):
+        a = np.arange(10)
+
+        assert_raises(IndexError, a.__getitem__, [0.5, 1.5])
+        assert_raises(IndexError, a.__getitem__, (['1', '2'],))
+
+        # The following is valid
+        a.__getitem__([])
+
+
+class TestMultipleEllipsisError:
+    """An index can only have a single ellipsis.
+
+    """
+    def test_basic(self):
+        a = np.arange(10)
+        assert_raises(IndexError, lambda: a[..., ...])
+        assert_raises(IndexError, a.__getitem__, ((Ellipsis,) * 2,))
+        assert_raises(IndexError, a.__getitem__, ((Ellipsis,) * 3,))
+
+
+class TestCApiAccess:
+    def test_getitem(self):
+        subscript = functools.partial(array_indexing, 0)
+
+        # 0-d arrays don't work:
+        assert_raises(IndexError, subscript, np.ones(()), 0)
+        # Out of bound values:
+        assert_raises(IndexError, subscript, np.ones(10), 11)
+        assert_raises(IndexError, subscript, np.ones(10), -11)
+        assert_raises(IndexError, subscript, np.ones((10, 10)), 11)
+        assert_raises(IndexError, subscript, np.ones((10, 10)), -11)
+
+        a = np.arange(10)
+        assert_array_equal(a[4], subscript(a, 4))
+        a = a.reshape(5, 2)
+        assert_array_equal(a[-4], subscript(a, -4))
+
+    def test_setitem(self):
+        assign = functools.partial(array_indexing, 1)
+
+        # Deletion is impossible:
+        assert_raises(ValueError, assign, np.ones(10), 0)
+        # 0-d arrays don't work:
+        assert_raises(IndexError, assign, np.ones(()), 0, 0)
+        # Out of bound values:
+        assert_raises(IndexError, assign, np.ones(10), 11, 0)
+        assert_raises(IndexError, assign, np.ones(10), -11, 0)
+        assert_raises(IndexError, assign, np.ones((10, 10)), 11, 0)
+        assert_raises(IndexError, assign, np.ones((10, 10)), -11, 0)
+
+        a = np.arange(10)
+        assign(a, 4, 10)
+        assert_(a[4] == 10)
+
+        a = a.reshape(5, 2)
+        assign(a, 4, 10)
+        assert_array_equal(a[-1], [10, 10])

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/core/tests/test_memmap.py

@@ -0,0 +1,215 @@
+import sys
+import os
+import mmap
+import pytest
+from pathlib import Path
+from tempfile import NamedTemporaryFile, TemporaryFile
+
+from numpy import (
+    memmap, sum, average, prod, ndarray, isscalar, add, subtract, multiply)
+
+from numpy import arange, allclose, asarray
+from numpy.testing import (
+    assert_, assert_equal, assert_array_equal, suppress_warnings, IS_PYPY,
+    break_cycles
+    )
+
+class TestMemmap:
+    def setup_method(self):
+        self.tmpfp = NamedTemporaryFile(prefix='mmap')
+        self.shape = (3, 4)
+        self.dtype = 'float32'
+        self.data = arange(12, dtype=self.dtype)
+        self.data.resize(self.shape)
+
+    def teardown_method(self):
+        self.tmpfp.close()
+        self.data = None
+        if IS_PYPY:
+            break_cycles()
+            break_cycles()
+
+    def test_roundtrip(self):
+        # Write data to file
+        fp = memmap(self.tmpfp, dtype=self.dtype, mode='w+',
+                    shape=self.shape)
+        fp[:] = self.data[:]
+        del fp  # Test __del__ machinery, which handles cleanup
+
+        # Read data back from file
+        newfp = memmap(self.tmpfp, dtype=self.dtype, mode='r',
+                       shape=self.shape)
+        assert_(allclose(self.data, newfp))
+        assert_array_equal(self.data, newfp)
+        assert_equal(newfp.flags.writeable, False)
+
+    def test_open_with_filename(self, tmp_path):
+        tmpname = tmp_path / 'mmap'
+        fp = memmap(tmpname, dtype=self.dtype, mode='w+',
+                       shape=self.shape)
+        fp[:] = self.data[:]
+        del fp
+
+    def test_unnamed_file(self):
+        with TemporaryFile() as f:
+            fp = memmap(f, dtype=self.dtype, shape=self.shape)
+            del fp
+
+    def test_attributes(self):
+        offset = 1
+        mode = "w+"
+        fp = memmap(self.tmpfp, dtype=self.dtype, mode=mode,
+                    shape=self.shape, offset=offset)
+        assert_equal(offset, fp.offset)
+        assert_equal(mode, fp.mode)
+        del fp
+
+    def test_filename(self, tmp_path):
+        tmpname = tmp_path / "mmap"
+        fp = memmap(tmpname, dtype=self.dtype, mode='w+',
+                       shape=self.shape)
+        abspath = Path(os.path.abspath(tmpname))
+        fp[:] = self.data[:]
+        assert_equal(abspath, fp.filename)
+        b = fp[:1]
+        assert_equal(abspath, b.filename)
+        del b
+        del fp
+
+    def test_path(self, tmp_path):
+        tmpname = tmp_path / "mmap"
+        fp = memmap(Path(tmpname), dtype=self.dtype, mode='w+',
+                       shape=self.shape)
+        # os.path.realpath does not resolve symlinks on Windows
+        # see: https://bugs.python.org/issue9949
+        # use Path.resolve, just as memmap class does internally
+        abspath = str(Path(tmpname).resolve())
+        fp[:] = self.data[:]
+        assert_equal(abspath, str(fp.filename.resolve()))
+        b = fp[:1]
+        assert_equal(abspath, str(b.filename.resolve()))
+        del b
+        del fp
+
+    def test_filename_fileobj(self):
+        fp = memmap(self.tmpfp, dtype=self.dtype, mode="w+",
+                    shape=self.shape)
+        assert_equal(fp.filename, self.tmpfp.name)
+
+    @pytest.mark.skipif(sys.platform == 'gnu0',
+                        reason="Known to fail on hurd")
+    def test_flush(self):
+        fp = memmap(self.tmpfp, dtype=self.dtype, mode='w+',
+                    shape=self.shape)
+        fp[:] = self.data[:]
+        assert_equal(fp[0], self.data[0])
+        fp.flush()
+
+    def test_del(self):
+        # Make sure a view does not delete the underlying mmap
+        fp_base = memmap(self.tmpfp, dtype=self.dtype, mode='w+',
+                    shape=self.shape)
+        fp_base[0] = 5
+        fp_view = fp_base[0:1]
+        assert_equal(fp_view[0], 5)
+        del fp_view
+        # Should still be able to access and assign values after
+        # deleting the view
+        assert_equal(fp_base[0], 5)
+        fp_base[0] = 6
+        assert_equal(fp_base[0], 6)
+
+    def test_arithmetic_drops_references(self):
+        fp = memmap(self.tmpfp, dtype=self.dtype, mode='w+',
+                    shape=self.shape)
+        tmp = (fp + 10)
+        if isinstance(tmp, memmap):
+            assert_(tmp._mmap is not fp._mmap)
+
+    def test_indexing_drops_references(self):
+        fp = memmap(self.tmpfp, dtype=self.dtype, mode='w+',
+                    shape=self.shape)
+        tmp = fp[(1, 2), (2, 3)]
+        if isinstance(tmp, memmap):
+            assert_(tmp._mmap is not fp._mmap)
+
+    def test_slicing_keeps_references(self):
+        fp = memmap(self.tmpfp, dtype=self.dtype, mode='w+',
+                    shape=self.shape)
+        assert_(fp[:2, :2]._mmap is fp._mmap)
+
+    def test_view(self):
+        fp = memmap(self.tmpfp, dtype=self.dtype, shape=self.shape)
+        new1 = fp.view()
+        new2 = new1.view()
+        assert_(new1.base is fp)
+        assert_(new2.base is fp)
+        new_array = asarray(fp)
+        assert_(new_array.base is fp)
+
+    def test_ufunc_return_ndarray(self):
+        fp = memmap(self.tmpfp, dtype=self.dtype, shape=self.shape)
+        fp[:] = self.data
+
+        with suppress_warnings() as sup:
+            sup.filter(FutureWarning, "np.average currently does not preserve")
+            for unary_op in [sum, average, prod]:
+                result = unary_op(fp)
+                assert_(isscalar(result))
+                assert_(result.__class__ is self.data[0, 0].__class__)
+
+                assert_(unary_op(fp, axis=0).__class__ is ndarray)
+                assert_(unary_op(fp, axis=1).__class__ is ndarray)
+
+        for binary_op in [add, subtract, multiply]:
+            assert_(binary_op(fp, self.data).__class__ is ndarray)
+            assert_(binary_op(self.data, fp).__class__ is ndarray)
+            assert_(binary_op(fp, fp).__class__ is ndarray)
+
+        fp += 1
+        assert(fp.__class__ is memmap)
+        add(fp, 1, out=fp)
+        assert(fp.__class__ is memmap)
+
+    def test_getitem(self):
+        fp = memmap(self.tmpfp, dtype=self.dtype, shape=self.shape)
+        fp[:] = self.data
+
+        assert_(fp[1:, :-1].__class__ is memmap)
+        # Fancy indexing returns a copy that is not memmapped
+        assert_(fp[[0, 1]].__class__ is ndarray)
+
+    def test_memmap_subclass(self):
+        class MemmapSubClass(memmap):
+            pass
+
+        fp = MemmapSubClass(self.tmpfp, dtype=self.dtype, shape=self.shape)
+        fp[:] = self.data
+
+        # We keep previous behavior for subclasses of memmap, i.e. the
+        # ufunc and __getitem__ output is never turned into a ndarray
+        assert_(sum(fp, axis=0).__class__ is MemmapSubClass)
+        assert_(sum(fp).__class__ is MemmapSubClass)
+        assert_(fp[1:, :-1].__class__ is MemmapSubClass)
+        assert(fp[[0, 1]].__class__ is MemmapSubClass)
+
+    def test_mmap_offset_greater_than_allocation_granularity(self):
+        size = 5 * mmap.ALLOCATIONGRANULARITY
+        offset = mmap.ALLOCATIONGRANULARITY + 1
+        fp = memmap(self.tmpfp, shape=size, mode='w+', offset=offset)
+        assert_(fp.offset == offset)
+
+    def test_no_shape(self):
+        self.tmpfp.write(b'a'*16)
+        mm = memmap(self.tmpfp, dtype='float64')
+        assert_equal(mm.shape, (2,))
+
+    def test_empty_array(self):
+        # gh-12653
+        with pytest.raises(ValueError, match='empty file'):
+            memmap(self.tmpfp, shape=(0,4), mode='w+')
+
+        self.tmpfp.write(b'\0')
+
+        # ok now the file is not empty
+        memmap(self.tmpfp, shape=(0,4), mode='w+')

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/core/tests/test_records.py

@@ -0,0 +1,520 @@
+import collections.abc
+import textwrap
+from io import BytesIO
+from os import path
+from pathlib import Path
+import pytest
+
+import numpy as np
+from numpy.testing import (
+    assert_, assert_equal, assert_array_equal, assert_array_almost_equal,
+    assert_raises, temppath,
+    )
+from numpy.compat import pickle
+
+
+class TestFromrecords:
+    def test_fromrecords(self):
+        r = np.rec.fromrecords([[456, 'dbe', 1.2], [2, 'de', 1.3]],
+                            names='col1,col2,col3')
+        assert_equal(r[0].item(), (456, 'dbe', 1.2))
+        assert_equal(r['col1'].dtype.kind, 'i')
+        assert_equal(r['col2'].dtype.kind, 'U')
+        assert_equal(r['col2'].dtype.itemsize, 12)
+        assert_equal(r['col3'].dtype.kind, 'f')
+
+    def test_fromrecords_0len(self):
+        """ Verify fromrecords works with a 0-length input """
+        dtype = [('a', float), ('b', float)]
+        r = np.rec.fromrecords([], dtype=dtype)
+        assert_equal(r.shape, (0,))
+
+    def test_fromrecords_2d(self):
+        data = [
+            [(1, 2), (3, 4), (5, 6)],
+            [(6, 5), (4, 3), (2, 1)]
+        ]
+        expected_a = [[1, 3, 5], [6, 4, 2]]
+        expected_b = [[2, 4, 6], [5, 3, 1]]
+
+        # try with dtype
+        r1 = np.rec.fromrecords(data, dtype=[('a', int), ('b', int)])
+        assert_equal(r1['a'], expected_a)
+        assert_equal(r1['b'], expected_b)
+
+        # try with names
+        r2 = np.rec.fromrecords(data, names=['a', 'b'])
+        assert_equal(r2['a'], expected_a)
+        assert_equal(r2['b'], expected_b)
+
+        assert_equal(r1, r2)
+
+    def test_method_array(self):
+        r = np.rec.array(b'abcdefg' * 100, formats='i2,a3,i4', shape=3, byteorder='big')
+        assert_equal(r[1].item(), (25444, b'efg', 1633837924))
+
+    def test_method_array2(self):
+        r = np.rec.array([(1, 11, 'a'), (2, 22, 'b'), (3, 33, 'c'), (4, 44, 'd'), (5, 55, 'ex'),
+                     (6, 66, 'f'), (7, 77, 'g')], formats='u1,f4,a1')
+        assert_equal(r[1].item(), (2, 22.0, b'b'))
+
+    def test_recarray_slices(self):
+        r = np.rec.array([(1, 11, 'a'), (2, 22, 'b'), (3, 33, 'c'), (4, 44, 'd'), (5, 55, 'ex'),
+                     (6, 66, 'f'), (7, 77, 'g')], formats='u1,f4,a1')
+        assert_equal(r[1::2][1].item(), (4, 44.0, b'd'))
+
+    def test_recarray_fromarrays(self):
+        x1 = np.array([1, 2, 3, 4])
+        x2 = np.array(['a', 'dd', 'xyz', '12'])
+        x3 = np.array([1.1, 2, 3, 4])
+        r = np.rec.fromarrays([x1, x2, x3], names='a,b,c')
+        assert_equal(r[1].item(), (2, 'dd', 2.0))
+        x1[1] = 34
+        assert_equal(r.a, np.array([1, 2, 3, 4]))
+
+    def test_recarray_fromfile(self):
+        data_dir = path.join(path.dirname(__file__), 'data')
+        filename = path.join(data_dir, 'recarray_from_file.fits')
+        fd = open(filename, 'rb')
+        fd.seek(2880 * 2)
+        r1 = np.rec.fromfile(fd, formats='f8,i4,a5', shape=3, byteorder='big')
+        fd.seek(2880 * 2)
+        r2 = np.rec.array(fd, formats='f8,i4,a5', shape=3, byteorder='big')
+        fd.seek(2880 * 2)
+        bytes_array = BytesIO()
+        bytes_array.write(fd.read())
+        bytes_array.seek(0)
+        r3 = np.rec.fromfile(bytes_array, formats='f8,i4,a5', shape=3, byteorder='big')
+        fd.close()
+        assert_equal(r1, r2)
+        assert_equal(r2, r3)
+
+    def test_recarray_from_obj(self):
+        count = 10
+        a = np.zeros(count, dtype='O')
+        b = np.zeros(count, dtype='f8')
+        c = np.zeros(count, dtype='f8')
+        for i in range(len(a)):
+            a[i] = list(range(1, 10))
+
+        mine = np.rec.fromarrays([a, b, c], names='date,data1,data2')
+        for i in range(len(a)):
+            assert_((mine.date[i] == list(range(1, 10))))
+            assert_((mine.data1[i] == 0.0))
+            assert_((mine.data2[i] == 0.0))
+
+    def test_recarray_repr(self):
+        a = np.array([(1, 0.1), (2, 0.2)],
+                     dtype=[('foo', '<i4'), ('bar', '<f8')])
+        a = np.rec.array(a)
+        assert_equal(
+            repr(a),
+            textwrap.dedent("""\
+            rec.array([(1, 0.1), (2, 0.2)],
+                      dtype=[('foo', '<i4'), ('bar', '<f8')])""")
+        )
+
+        # make sure non-structured dtypes also show up as rec.array
+        a = np.array(np.ones(4, dtype='f8'))
+        assert_(repr(np.rec.array(a)).startswith('rec.array'))
+
+        # check that the 'np.record' part of the dtype isn't shown
+        a = np.rec.array(np.ones(3, dtype='i4,i4'))
+        assert_equal(repr(a).find('numpy.record'), -1)
+        a = np.rec.array(np.ones(3, dtype='i4'))
+        assert_(repr(a).find('dtype=int32') != -1)
+
+    def test_0d_recarray_repr(self):
+        arr_0d = np.rec.array((1, 2.0, '2003'), dtype='<i4,<f8,<M8[Y]')
+        assert_equal(repr(arr_0d), textwrap.dedent("""\
+            rec.array((1, 2., '2003'),
+                      dtype=[('f0', '<i4'), ('f1', '<f8'), ('f2', '<M8[Y]')])"""))
+
+        record = arr_0d[()]
+        assert_equal(repr(record), "(1, 2., '2003')")
+        # 1.13 converted to python scalars before the repr
+        try:
+            np.set_printoptions(legacy='1.13')
+            assert_equal(repr(record), '(1, 2.0, datetime.date(2003, 1, 1))')
+        finally:
+            np.set_printoptions(legacy=False)
+
+    def test_recarray_from_repr(self):
+        a = np.array([(1,'ABC'), (2, "DEF")],
+                     dtype=[('foo', int), ('bar', 'S4')])
+        recordarr = np.rec.array(a)
+        recarr = a.view(np.recarray)
+        recordview = a.view(np.dtype((np.record, a.dtype)))
+
+        recordarr_r = eval("numpy." + repr(recordarr), {'numpy': np})
+        recarr_r = eval("numpy." + repr(recarr), {'numpy': np})
+        recordview_r = eval("numpy." + repr(recordview), {'numpy': np})
+
+        assert_equal(type(recordarr_r), np.recarray)
+        assert_equal(recordarr_r.dtype.type, np.record)
+        assert_equal(recordarr, recordarr_r)
+
+        assert_equal(type(recarr_r), np.recarray)
+        assert_equal(recarr_r.dtype.type, np.record)
+        assert_equal(recarr, recarr_r)
+
+        assert_equal(type(recordview_r), np.ndarray)
+        assert_equal(recordview.dtype.type, np.record)
+        assert_equal(recordview, recordview_r)
+
+    def test_recarray_views(self):
+        a = np.array([(1,'ABC'), (2, "DEF")],
+                     dtype=[('foo', int), ('bar', 'S4')])
+        b = np.array([1,2,3,4,5], dtype=np.int64)
+
+        #check that np.rec.array gives right dtypes
+        assert_equal(np.rec.array(a).dtype.type, np.record)
+        assert_equal(type(np.rec.array(a)), np.recarray)
+        assert_equal(np.rec.array(b).dtype.type, np.int64)
+        assert_equal(type(np.rec.array(b)), np.recarray)
+
+        #check that viewing as recarray does the same
+        assert_equal(a.view(np.recarray).dtype.type, np.record)
+        assert_equal(type(a.view(np.recarray)), np.recarray)
+        assert_equal(b.view(np.recarray).dtype.type, np.int64)
+        assert_equal(type(b.view(np.recarray)), np.recarray)
+
+        #check that view to non-structured dtype preserves type=np.recarray
+        r = np.rec.array(np.ones(4, dtype="f4,i4"))
+        rv = r.view('f8').view('f4,i4')
+        assert_equal(type(rv), np.recarray)
+        assert_equal(rv.dtype.type, np.record)
+
+        #check that getitem also preserves np.recarray and np.record
+        r = np.rec.array(np.ones(4, dtype=[('a', 'i4'), ('b', 'i4'),
+                                           ('c', 'i4,i4')]))
+        assert_equal(r['c'].dtype.type, np.record)
+        assert_equal(type(r['c']), np.recarray)
+
+        #and that it preserves subclasses (gh-6949)
+        class C(np.recarray):
+            pass
+
+        c = r.view(C)
+        assert_equal(type(c['c']), C)
+
+        # check that accessing nested structures keep record type, but
+        # not for subarrays, non-void structures, non-structured voids
+        test_dtype = [('a', 'f4,f4'), ('b', 'V8'), ('c', ('f4',2)),
+                      ('d', ('i8', 'i4,i4'))]
+        r = np.rec.array([((1,1), b'11111111', [1,1], 1),
+                          ((1,1), b'11111111', [1,1], 1)], dtype=test_dtype)
+        assert_equal(r.a.dtype.type, np.record)
+        assert_equal(r.b.dtype.type, np.void)
+        assert_equal(r.c.dtype.type, np.float32)
+        assert_equal(r.d.dtype.type, np.int64)
+        # check the same, but for views
+        r = np.rec.array(np.ones(4, dtype='i4,i4'))
+        assert_equal(r.view('f4,f4').dtype.type, np.record)
+        assert_equal(r.view(('i4',2)).dtype.type, np.int32)
+        assert_equal(r.view('V8').dtype.type, np.void)
+        assert_equal(r.view(('i8', 'i4,i4')).dtype.type, np.int64)
+
+        #check that we can undo the view
+        arrs = [np.ones(4, dtype='f4,i4'), np.ones(4, dtype='f8')]
+        for arr in arrs:
+            rec = np.rec.array(arr)
+            # recommended way to view as an ndarray:
+            arr2 = rec.view(rec.dtype.fields or rec.dtype, np.ndarray)
+            assert_equal(arr2.dtype.type, arr.dtype.type)
+            assert_equal(type(arr2), type(arr))
+
+    def test_recarray_from_names(self):
+        ra = np.rec.array([
+            (1, 'abc', 3.7000002861022949, 0),
+            (2, 'xy', 6.6999998092651367, 1),
+            (0, ' ', 0.40000000596046448, 0)],
+                       names='c1, c2, c3, c4')
+        pa = np.rec.fromrecords([
+            (1, 'abc', 3.7000002861022949, 0),
+            (2, 'xy', 6.6999998092651367, 1),
+            (0, ' ', 0.40000000596046448, 0)],
+                       names='c1, c2, c3, c4')
+        assert_(ra.dtype == pa.dtype)
+        assert_(ra.shape == pa.shape)
+        for k in range(len(ra)):
+            assert_(ra[k].item() == pa[k].item())
+
+    def test_recarray_conflict_fields(self):
+        ra = np.rec.array([(1, 'abc', 2.3), (2, 'xyz', 4.2),
+                        (3, 'wrs', 1.3)],
+                       names='field, shape, mean')
+        ra.mean = [1.1, 2.2, 3.3]
+        assert_array_almost_equal(ra['mean'], [1.1, 2.2, 3.3])
+        assert_(type(ra.mean) is type(ra.var))
+        ra.shape = (1, 3)
+        assert_(ra.shape == (1, 3))
+        ra.shape = ['A', 'B', 'C']
+        assert_array_equal(ra['shape'], [['A', 'B', 'C']])
+        ra.field = 5
+        assert_array_equal(ra['field'], [[5, 5, 5]])
+        assert_(isinstance(ra.field, collections.abc.Callable))
+
+    def test_fromrecords_with_explicit_dtype(self):
+        a = np.rec.fromrecords([(1, 'a'), (2, 'bbb')],
+                                dtype=[('a', int), ('b', object)])
+        assert_equal(a.a, [1, 2])
+        assert_equal(a[0].a, 1)
+        assert_equal(a.b, ['a', 'bbb'])
+        assert_equal(a[-1].b, 'bbb')
+        #
+        ndtype = np.dtype([('a', int), ('b', object)])
+        a = np.rec.fromrecords([(1, 'a'), (2, 'bbb')], dtype=ndtype)
+        assert_equal(a.a, [1, 2])
+        assert_equal(a[0].a, 1)
+        assert_equal(a.b, ['a', 'bbb'])
+        assert_equal(a[-1].b, 'bbb')
+
+    def test_recarray_stringtypes(self):
+        # Issue #3993
+        a = np.array([('abc ', 1), ('abc', 2)],
+                     dtype=[('foo', 'S4'), ('bar', int)])
+        a = a.view(np.recarray)
+        assert_equal(a.foo[0] == a.foo[1], False)
+
+    def test_recarray_returntypes(self):
+        qux_fields = {'C': (np.dtype('S5'), 0), 'D': (np.dtype('S5'), 6)}
+        a = np.rec.array([('abc ', (1,1), 1, ('abcde', 'fgehi')),
+                          ('abc', (2,3), 1, ('abcde', 'jklmn'))],
+                         dtype=[('foo', 'S4'),
+                                ('bar', [('A', int), ('B', int)]),
+                                ('baz', int), ('qux', qux_fields)])
+        assert_equal(type(a.foo), np.ndarray)
+        assert_equal(type(a['foo']), np.ndarray)
+        assert_equal(type(a.bar), np.recarray)
+        assert_equal(type(a['bar']), np.recarray)
+        assert_equal(a.bar.dtype.type, np.record)
+        assert_equal(type(a['qux']), np.recarray)
+        assert_equal(a.qux.dtype.type, np.record)
+        assert_equal(dict(a.qux.dtype.fields), qux_fields)
+        assert_equal(type(a.baz), np.ndarray)
+        assert_equal(type(a['baz']), np.ndarray)
+        assert_equal(type(a[0].bar), np.record)
+        assert_equal(type(a[0]['bar']), np.record)
+        assert_equal(a[0].bar.A, 1)
+        assert_equal(a[0].bar['A'], 1)
+        assert_equal(a[0]['bar'].A, 1)
+        assert_equal(a[0]['bar']['A'], 1)
+        assert_equal(a[0].qux.D, b'fgehi')
+        assert_equal(a[0].qux['D'], b'fgehi')
+        assert_equal(a[0]['qux'].D, b'fgehi')
+        assert_equal(a[0]['qux']['D'], b'fgehi')
+
+    def test_zero_width_strings(self):
+        # Test for #6430, based on the test case from #1901
+
+        cols = [['test'] * 3, [''] * 3]
+        rec = np.rec.fromarrays(cols)
+        assert_equal(rec['f0'], ['test', 'test', 'test'])
+        assert_equal(rec['f1'], ['', '', ''])
+
+        dt = np.dtype([('f0', '|S4'), ('f1', '|S')])
+        rec = np.rec.fromarrays(cols, dtype=dt)
+        assert_equal(rec.itemsize, 4)
+        assert_equal(rec['f0'], [b'test', b'test', b'test'])
+        assert_equal(rec['f1'], [b'', b'', b''])
+
+
+class TestPathUsage:
+    # Test that pathlib.Path can be used
+    def test_tofile_fromfile(self):
+        with temppath(suffix='.bin') as path:
+            path = Path(path)
+            np.random.seed(123)
+            a = np.random.rand(10).astype('f8,i4,a5')
+            a[5] = (0.5,10,'abcde')
+            with path.open("wb") as fd:
+                a.tofile(fd)
+            x = np.core.records.fromfile(path,
+                                         formats='f8,i4,a5',
+                                         shape=10)
+            assert_array_equal(x, a)
+
+
+class TestRecord:
+    def setup_method(self):
+        self.data = np.rec.fromrecords([(1, 2, 3), (4, 5, 6)],
+                            dtype=[("col1", "<i4"),
+                                   ("col2", "<i4"),
+                                   ("col3", "<i4")])
+
+    def test_assignment1(self):
+        a = self.data
+        assert_equal(a.col1[0], 1)
+        a[0].col1 = 0
+        assert_equal(a.col1[0], 0)
+
+    def test_assignment2(self):
+        a = self.data
+        assert_equal(a.col1[0], 1)
+        a.col1[0] = 0
+        assert_equal(a.col1[0], 0)
+
+    def test_invalid_assignment(self):
+        a = self.data
+
+        def assign_invalid_column(x):
+            x[0].col5 = 1
+
+        assert_raises(AttributeError, assign_invalid_column, a)
+
+    def test_nonwriteable_setfield(self):
+        # gh-8171
+        r = np.rec.array([(0,), (1,)], dtype=[('f', 'i4')])
+        r.flags.writeable = False
+        with assert_raises(ValueError):
+            r.f = [2, 3]
+        with assert_raises(ValueError):
+            r.setfield([2,3], *r.dtype.fields['f'])
+
+    def test_out_of_order_fields(self):
+        # names in the same order, padding added to descr
+        x = self.data[['col1', 'col2']]
+        assert_equal(x.dtype.names, ('col1', 'col2'))
+        assert_equal(x.dtype.descr,
+                     [('col1', '<i4'), ('col2', '<i4'), ('', '|V4')])
+
+        # names change order to match indexing, as of 1.14 - descr can't
+        # represent that
+        y = self.data[['col2', 'col1']]
+        assert_equal(y.dtype.names, ('col2', 'col1'))
+        assert_raises(ValueError, lambda: y.dtype.descr)
+
+    def test_pickle_1(self):
+        # Issue #1529
+        a = np.array([(1, [])], dtype=[('a', np.int32), ('b', np.int32, 0)])
+        for proto in range(2, pickle.HIGHEST_PROTOCOL + 1):
+            assert_equal(a, pickle.loads(pickle.dumps(a, protocol=proto)))
+            assert_equal(a[0], pickle.loads(pickle.dumps(a[0],
+                                                         protocol=proto)))
+
+    def test_pickle_2(self):
+        a = self.data
+        for proto in range(2, pickle.HIGHEST_PROTOCOL + 1):
+            assert_equal(a, pickle.loads(pickle.dumps(a, protocol=proto)))
+            assert_equal(a[0], pickle.loads(pickle.dumps(a[0],
+                                                         protocol=proto)))
+
+    def test_pickle_3(self):
+        # Issue #7140
+        a = self.data
+        for proto in range(2, pickle.HIGHEST_PROTOCOL + 1):
+            pa = pickle.loads(pickle.dumps(a[0], protocol=proto))
+            assert_(pa.flags.c_contiguous)
+            assert_(pa.flags.f_contiguous)
+            assert_(pa.flags.writeable)
+            assert_(pa.flags.aligned)
+
+    def test_pickle_void(self):
+        # issue gh-13593
+        dt = np.dtype([('obj', 'O'), ('int', 'i')])
+        a = np.empty(1, dtype=dt)
+        data = (bytearray(b'eman'),)
+        a['obj'] = data
+        a['int'] = 42
+        ctor, args = a[0].__reduce__()
+        # check the constructor is what we expect before interpreting the arguments
+        assert ctor is np.core.multiarray.scalar
+        dtype, obj = args
+        # make sure we did not pickle the address
+        assert not isinstance(obj, bytes)
+
+        assert_raises(RuntimeError, ctor, dtype, 13)
+
+        # Test roundtrip:
+        dump = pickle.dumps(a[0])
+        unpickled = pickle.loads(dump)
+        assert a[0] == unpickled
+
+        # Also check the similar (impossible) "object scalar" path:
+        with pytest.warns(DeprecationWarning):
+            assert ctor(np.dtype("O"), data) is data
+
+    def test_objview_record(self):
+        # https://github.com/numpy/numpy/issues/2599
+        dt = np.dtype([('foo', 'i8'), ('bar', 'O')])
+        r = np.zeros((1,3), dtype=dt).view(np.recarray)
+        r.foo = np.array([1, 2, 3])  # TypeError?
+
+        # https://github.com/numpy/numpy/issues/3256
+        ra = np.recarray((2,), dtype=[('x', object), ('y', float), ('z', int)])
+        ra[['x','y']]  # TypeError?
+
+    def test_record_scalar_setitem(self):
+        # https://github.com/numpy/numpy/issues/3561
+        rec = np.recarray(1, dtype=[('x', float, 5)])
+        rec[0].x = 1
+        assert_equal(rec[0].x, np.ones(5))
+
+    def test_missing_field(self):
+        # https://github.com/numpy/numpy/issues/4806
+        arr = np.zeros((3,), dtype=[('x', int), ('y', int)])
+        assert_raises(KeyError, lambda: arr[['nofield']])
+
+    def test_fromarrays_nested_structured_arrays(self):
+        arrays = [
+            np.arange(10),
+            np.ones(10, dtype=[('a', '<u2'), ('b', '<f4')]),
+        ]
+        arr = np.rec.fromarrays(arrays)  # ValueError?
+
+    @pytest.mark.parametrize('nfields', [0, 1, 2])
+    def test_assign_dtype_attribute(self, nfields):
+        dt = np.dtype([('a', np.uint8), ('b', np.uint8), ('c', np.uint8)][:nfields])
+        data = np.zeros(3, dt).view(np.recarray)
+
+        # the original and resulting dtypes differ on whether they are records
+        assert data.dtype.type == np.record
+        assert dt.type != np.record
+
+        # ensure that the dtype remains a record even when assigned
+        data.dtype = dt
+        assert data.dtype.type == np.record
+
+    @pytest.mark.parametrize('nfields', [0, 1, 2])
+    def test_nested_fields_are_records(self, nfields):
+        """ Test that nested structured types are treated as records too """
+        dt = np.dtype([('a', np.uint8), ('b', np.uint8), ('c', np.uint8)][:nfields])
+        dt_outer = np.dtype([('inner', dt)])
+
+        data = np.zeros(3, dt_outer).view(np.recarray)
+        assert isinstance(data, np.recarray)
+        assert isinstance(data['inner'], np.recarray)
+
+        data0 = data[0]
+        assert isinstance(data0, np.record)
+        assert isinstance(data0['inner'], np.record)
+
+    def test_nested_dtype_padding(self):
+        """ test that trailing padding is preserved """
+        # construct a dtype with padding at the end
+        dt = np.dtype([('a', np.uint8), ('b', np.uint8), ('c', np.uint8)])
+        dt_padded_end = dt[['a', 'b']]
+        assert dt_padded_end.itemsize == dt.itemsize
+
+        dt_outer = np.dtype([('inner', dt_padded_end)])
+
+        data = np.zeros(3, dt_outer).view(np.recarray)
+        assert_equal(data['inner'].dtype, dt_padded_end)
+
+        data0 = data[0]
+        assert_equal(data0['inner'].dtype, dt_padded_end)
+
+
+def test_find_duplicate():
+    l1 = [1, 2, 3, 4, 5, 6]
+    assert_(np.rec.find_duplicate(l1) == [])
+
+    l2 = [1, 2, 1, 4, 5, 6]
+    assert_(np.rec.find_duplicate(l2) == [1])
+
+    l3 = [1, 2, 1, 4, 1, 6, 2, 3]
+    assert_(np.rec.find_duplicate(l3) == [1, 2])
+
+    l3 = [2, 2, 1, 4, 1, 6, 2, 3]
+    assert_(np.rec.find_duplicate(l3) == [2, 1])

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/core/tests/test_scalar_methods.py

@@ -0,0 +1,204 @@
+"""
+Test the scalar constructors, which also do type-coercion
+"""
+import fractions
+import platform
+import types
+from typing import Any, Type
+
+import pytest
+import numpy as np
+
+from numpy.testing import assert_equal, assert_raises, IS_MUSL
+
+
+class TestAsIntegerRatio:
+    # derived in part from the cpython test "test_floatasratio"
+
+    @pytest.mark.parametrize("ftype", [
+        np.half, np.single, np.double, np.longdouble])
+    @pytest.mark.parametrize("f, ratio", [
+        (0.875, (7, 8)),
+        (-0.875, (-7, 8)),
+        (0.0, (0, 1)),
+        (11.5, (23, 2)),
+        ])
+    def test_small(self, ftype, f, ratio):
+        assert_equal(ftype(f).as_integer_ratio(), ratio)
+
+    @pytest.mark.parametrize("ftype", [
+        np.half, np.single, np.double, np.longdouble])
+    def test_simple_fractions(self, ftype):
+        R = fractions.Fraction
+        assert_equal(R(0, 1),
+                     R(*ftype(0.0).as_integer_ratio()))
+        assert_equal(R(5, 2),
+                     R(*ftype(2.5).as_integer_ratio()))
+        assert_equal(R(1, 2),
+                     R(*ftype(0.5).as_integer_ratio()))
+        assert_equal(R(-2100, 1),
+                     R(*ftype(-2100.0).as_integer_ratio()))
+
+    @pytest.mark.parametrize("ftype", [
+        np.half, np.single, np.double, np.longdouble])
+    def test_errors(self, ftype):
+        assert_raises(OverflowError, ftype('inf').as_integer_ratio)
+        assert_raises(OverflowError, ftype('-inf').as_integer_ratio)
+        assert_raises(ValueError, ftype('nan').as_integer_ratio)
+
+    def test_against_known_values(self):
+        R = fractions.Fraction
+        assert_equal(R(1075, 512),
+                     R(*np.half(2.1).as_integer_ratio()))
+        assert_equal(R(-1075, 512),
+                     R(*np.half(-2.1).as_integer_ratio()))
+        assert_equal(R(4404019, 2097152),
+                     R(*np.single(2.1).as_integer_ratio()))
+        assert_equal(R(-4404019, 2097152),
+                     R(*np.single(-2.1).as_integer_ratio()))
+        assert_equal(R(4728779608739021, 2251799813685248),
+                     R(*np.double(2.1).as_integer_ratio()))
+        assert_equal(R(-4728779608739021, 2251799813685248),
+                     R(*np.double(-2.1).as_integer_ratio()))
+        # longdouble is platform dependent
+
+    @pytest.mark.parametrize("ftype, frac_vals, exp_vals", [
+        # dtype test cases generated using hypothesis
+        # first five generated cases per dtype
+        (np.half, [0.0, 0.01154830649280303, 0.31082276347447274,
+                   0.527350517124794, 0.8308562335072596],
+                  [0, 1, 0, -8, 12]),
+        (np.single, [0.0, 0.09248576989263226, 0.8160498218131407,
+                     0.17389442853722373, 0.7956044195067877],
+                    [0, 12, 10, 17, -26]),
+        (np.double, [0.0, 0.031066908499895136, 0.5214135908877832,
+                     0.45780736035689296, 0.5906586745934036],
+                    [0, -801, 51, 194, -653]),
+        pytest.param(
+            np.longdouble,
+            [0.0, 0.20492557202724854, 0.4277180662199366, 0.9888085019891495,
+             0.9620175814461964],
+            [0, -7400, 14266, -7822, -8721],
+            marks=[
+                pytest.mark.skipif(
+                    np.finfo(np.double) == np.finfo(np.longdouble),
+                    reason="long double is same as double"),
+                pytest.mark.skipif(
+                    platform.machine().startswith("ppc"),
+                    reason="IBM double double"),
+            ]
+        )
+    ])
+    def test_roundtrip(self, ftype, frac_vals, exp_vals):
+        for frac, exp in zip(frac_vals, exp_vals):
+            f = np.ldexp(ftype(frac), exp)
+            assert f.dtype == ftype
+            n, d = f.as_integer_ratio()
+
+            try:
+                nf = np.longdouble(n)
+                df = np.longdouble(d)
+                if not np.isfinite(df):
+                    raise OverflowError
+            except (OverflowError, RuntimeWarning):
+                # the values may not fit in any float type
+                pytest.skip("longdouble too small on this platform")
+
+            assert_equal(nf / df, f, "{}/{}".format(n, d))
+
+
+class TestIsInteger:
+    @pytest.mark.parametrize("str_value", ["inf", "nan"])
+    @pytest.mark.parametrize("code", np.typecodes["Float"])
+    def test_special(self, code: str, str_value: str) -> None:
+        cls = np.dtype(code).type
+        value = cls(str_value)
+        assert not value.is_integer()
+
+    @pytest.mark.parametrize(
+        "code", np.typecodes["Float"] + np.typecodes["AllInteger"]
+    )
+    def test_true(self, code: str) -> None:
+        float_array = np.arange(-5, 5).astype(code)
+        for value in float_array:
+            assert value.is_integer()
+
+    @pytest.mark.parametrize("code", np.typecodes["Float"])
+    def test_false(self, code: str) -> None:
+        float_array = np.arange(-5, 5).astype(code)
+        float_array *= 1.1
+        for value in float_array:
+            if value == 0:
+                continue
+            assert not value.is_integer()
+
+
+class TestClassGetItem:
+    @pytest.mark.parametrize("cls", [
+        np.number,
+        np.integer,
+        np.inexact,
+        np.unsignedinteger,
+        np.signedinteger,
+        np.floating,
+    ])
+    def test_abc(self, cls: Type[np.number]) -> None:
+        alias = cls[Any]
+        assert isinstance(alias, types.GenericAlias)
+        assert alias.__origin__ is cls
+
+    def test_abc_complexfloating(self) -> None:
+        alias = np.complexfloating[Any, Any]
+        assert isinstance(alias, types.GenericAlias)
+        assert alias.__origin__ is np.complexfloating
+
+    @pytest.mark.parametrize("arg_len", range(4))
+    def test_abc_complexfloating_subscript_tuple(self, arg_len: int) -> None:
+        arg_tup = (Any,) * arg_len
+        if arg_len in (1, 2):
+            assert np.complexfloating[arg_tup]
+        else:
+            match = f"Too {'few' if arg_len == 0 else 'many'} arguments"
+            with pytest.raises(TypeError, match=match):
+                np.complexfloating[arg_tup]
+
+    @pytest.mark.parametrize("cls", [np.generic, np.flexible, np.character])
+    def test_abc_non_numeric(self, cls: Type[np.generic]) -> None:
+        with pytest.raises(TypeError):
+            cls[Any]
+
+    @pytest.mark.parametrize("code", np.typecodes["All"])
+    def test_concrete(self, code: str) -> None:
+        cls = np.dtype(code).type
+        with pytest.raises(TypeError):
+            cls[Any]
+
+    @pytest.mark.parametrize("arg_len", range(4))
+    def test_subscript_tuple(self, arg_len: int) -> None:
+        arg_tup = (Any,) * arg_len
+        if arg_len == 1:
+            assert np.number[arg_tup]
+        else:
+            with pytest.raises(TypeError):
+                np.number[arg_tup]
+
+    def test_subscript_scalar(self) -> None:
+        assert np.number[Any]
+
+
+class TestBitCount:
+    # derived in part from the cpython test "test_bit_count"
+
+    @pytest.mark.parametrize("itype", np.sctypes['int']+np.sctypes['uint'])
+    def test_small(self, itype):
+        for a in range(max(np.iinfo(itype).min, 0), 128):
+            msg = f"Smoke test for {itype}({a}).bit_count()"
+            assert itype(a).bit_count() == bin(a).count("1"), msg
+
+    def test_bit_count(self):
+        for exp in [10, 17, 63]:
+            a = 2**exp
+            assert np.uint64(a).bit_count() == 1
+            assert np.uint64(a - 1).bit_count() == exp
+            assert np.uint64(a ^ 63).bit_count() == 7
+            assert np.uint64((a - 1) ^ 510).bit_count() == exp - 8

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/core/tests/test_scalarbuffer.py

@@ -0,0 +1,153 @@
+"""
+Test scalar buffer interface adheres to PEP 3118
+"""
+import numpy as np
+from numpy.core._rational_tests import rational
+from numpy.core._multiarray_tests import get_buffer_info
+import pytest
+
+from numpy.testing import assert_, assert_equal, assert_raises
+
+# PEP3118 format strings for native (standard alignment and byteorder) types
+scalars_and_codes = [
+    (np.bool_, '?'),
+    (np.byte, 'b'),
+    (np.short, 'h'),
+    (np.intc, 'i'),
+    (np.int_, 'l'),
+    (np.longlong, 'q'),
+    (np.ubyte, 'B'),
+    (np.ushort, 'H'),
+    (np.uintc, 'I'),
+    (np.uint, 'L'),
+    (np.ulonglong, 'Q'),
+    (np.half, 'e'),
+    (np.single, 'f'),
+    (np.double, 'd'),
+    (np.longdouble, 'g'),
+    (np.csingle, 'Zf'),
+    (np.cdouble, 'Zd'),
+    (np.clongdouble, 'Zg'),
+]
+scalars_only, codes_only = zip(*scalars_and_codes)
+
+
+class TestScalarPEP3118:
+
+    @pytest.mark.parametrize('scalar', scalars_only, ids=codes_only)
+    def test_scalar_match_array(self, scalar):
+        x = scalar()
+        a = np.array([], dtype=np.dtype(scalar))
+        mv_x = memoryview(x)
+        mv_a = memoryview(a)
+        assert_equal(mv_x.format, mv_a.format)
+
+    @pytest.mark.parametrize('scalar', scalars_only, ids=codes_only)
+    def test_scalar_dim(self, scalar):
+        x = scalar()
+        mv_x = memoryview(x)
+        assert_equal(mv_x.itemsize, np.dtype(scalar).itemsize)
+        assert_equal(mv_x.ndim, 0)
+        assert_equal(mv_x.shape, ())
+        assert_equal(mv_x.strides, ())
+        assert_equal(mv_x.suboffsets, ())
+
+    @pytest.mark.parametrize('scalar, code', scalars_and_codes, ids=codes_only)
+    def test_scalar_code_and_properties(self, scalar, code):
+        x = scalar()
+        expected = dict(strides=(), itemsize=x.dtype.itemsize, ndim=0,
+                        shape=(), format=code, readonly=True)
+
+        mv_x = memoryview(x)
+        assert self._as_dict(mv_x) == expected
+
+    @pytest.mark.parametrize('scalar', scalars_only, ids=codes_only)
+    def test_scalar_buffers_readonly(self, scalar):
+        x = scalar()
+        with pytest.raises(BufferError, match="scalar buffer is readonly"):
+            get_buffer_info(x, ["WRITABLE"])
+
+    def test_void_scalar_structured_data(self):
+        dt = np.dtype([('name', np.str_, 16), ('grades', np.float64, (2,))])
+        x = np.array(('ndarray_scalar', (1.2, 3.0)), dtype=dt)[()]
+        assert_(isinstance(x, np.void))
+        mv_x = memoryview(x)
+        expected_size = 16 * np.dtype((np.str_, 1)).itemsize
+        expected_size += 2 * np.dtype(np.float64).itemsize
+        assert_equal(mv_x.itemsize, expected_size)
+        assert_equal(mv_x.ndim, 0)
+        assert_equal(mv_x.shape, ())
+        assert_equal(mv_x.strides, ())
+        assert_equal(mv_x.suboffsets, ())
+
+        # check scalar format string against ndarray format string
+        a = np.array([('Sarah', (8.0, 7.0)), ('John', (6.0, 7.0))], dtype=dt)
+        assert_(isinstance(a, np.ndarray))
+        mv_a = memoryview(a)
+        assert_equal(mv_x.itemsize, mv_a.itemsize)
+        assert_equal(mv_x.format, mv_a.format)
+
+        # Check that we do not allow writeable buffer export (technically
+        # we could allow it sometimes here...)
+        with pytest.raises(BufferError, match="scalar buffer is readonly"):
+            get_buffer_info(x, ["WRITABLE"])
+
+    def _as_dict(self, m):
+        return dict(strides=m.strides, shape=m.shape, itemsize=m.itemsize,
+                    ndim=m.ndim, format=m.format, readonly=m.readonly)
+
+    def test_datetime_memoryview(self):
+        # gh-11656
+        # Values verified with v1.13.3, shape is not () as in test_scalar_dim
+
+        dt1 = np.datetime64('2016-01-01')
+        dt2 = np.datetime64('2017-01-01')
+        expected = dict(strides=(1,), itemsize=1, ndim=1, shape=(8,),
+                        format='B', readonly=True)
+        v = memoryview(dt1)
+        assert self._as_dict(v) == expected
+
+        v = memoryview(dt2 - dt1)
+        assert self._as_dict(v) == expected
+
+        dt = np.dtype([('a', 'uint16'), ('b', 'M8[s]')])
+        a = np.empty(1, dt)
+        # Fails to create a PEP 3118 valid buffer
+        assert_raises((ValueError, BufferError), memoryview, a[0])
+
+        # Check that we do not allow writeable buffer export
+        with pytest.raises(BufferError, match="scalar buffer is readonly"):
+            get_buffer_info(dt1, ["WRITABLE"])
+
+    @pytest.mark.parametrize('s', [
+        pytest.param("\x32\x32", id="ascii"),
+        pytest.param("\uFE0F\uFE0F", id="basic multilingual"),
+        pytest.param("\U0001f4bb\U0001f4bb", id="non-BMP"),
+    ])
+    def test_str_ucs4(self, s):
+        s = np.str_(s)  # only our subclass implements the buffer protocol
+
+        # all the same, characters always encode as ucs4
+        expected = dict(strides=(), itemsize=8, ndim=0, shape=(), format='2w',
+                        readonly=True)
+
+        v = memoryview(s)
+        assert self._as_dict(v) == expected
+
+        # integers of the paltform-appropriate endianness
+        code_points = np.frombuffer(v, dtype='i4')
+
+        assert_equal(code_points, [ord(c) for c in s])
+
+        # Check that we do not allow writeable buffer export
+        with pytest.raises(BufferError, match="scalar buffer is readonly"):
+            get_buffer_info(s, ["WRITABLE"])
+
+    def test_user_scalar_fails_buffer(self):
+        r = rational(1)
+        with assert_raises(TypeError):
+            memoryview(r)
+
+        # Check that we do not allow writeable buffer export
+        with pytest.raises(BufferError, match="scalar buffer is readonly"):
+            get_buffer_info(r, ["WRITABLE"])

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/core/tests/test_scalarinherit.py

@@ -0,0 +1,98 @@
+""" Test printing of scalar types.
+
+"""
+import pytest
+
+import numpy as np
+from numpy.testing import assert_, assert_raises
+
+
+class A:
+    pass
+class B(A, np.float64):
+    pass
+
+class C(B):
+    pass
+class D(C, B):
+    pass
+
+class B0(np.float64, A):
+    pass
+class C0(B0):
+    pass
+
+class HasNew:
+    def __new__(cls, *args, **kwargs):
+        return cls, args, kwargs
+
+class B1(np.float64, HasNew):
+    pass
+
+
+class TestInherit:
+    def test_init(self):
+        x = B(1.0)
+        assert_(str(x) == '1.0')
+        y = C(2.0)
+        assert_(str(y) == '2.0')
+        z = D(3.0)
+        assert_(str(z) == '3.0')
+
+    def test_init2(self):
+        x = B0(1.0)
+        assert_(str(x) == '1.0')
+        y = C0(2.0)
+        assert_(str(y) == '2.0')
+
+    def test_gh_15395(self):
+        # HasNew is the second base, so `np.float64` should have priority
+        x = B1(1.0)
+        assert_(str(x) == '1.0')
+
+        # previously caused RecursionError!?
+        with pytest.raises(TypeError):
+            B1(1.0, 2.0)
+
+
+class TestCharacter:
+    def test_char_radd(self):
+        # GH issue 9620, reached gentype_add and raise TypeError
+        np_s = np.bytes_('abc')
+        np_u = np.str_('abc')
+        s = b'def'
+        u = 'def'
+        assert_(np_s.__radd__(np_s) is NotImplemented)
+        assert_(np_s.__radd__(np_u) is NotImplemented)
+        assert_(np_s.__radd__(s) is NotImplemented)
+        assert_(np_s.__radd__(u) is NotImplemented)
+        assert_(np_u.__radd__(np_s) is NotImplemented)
+        assert_(np_u.__radd__(np_u) is NotImplemented)
+        assert_(np_u.__radd__(s) is NotImplemented)
+        assert_(np_u.__radd__(u) is NotImplemented)
+        assert_(s + np_s == b'defabc')
+        assert_(u + np_u == 'defabc')
+
+        class MyStr(str, np.generic):
+            # would segfault
+            pass
+
+        with assert_raises(TypeError):
+            # Previously worked, but gave completely wrong result
+            ret = s + MyStr('abc')
+
+        class MyBytes(bytes, np.generic):
+            # would segfault
+            pass
+
+        ret = s + MyBytes(b'abc')
+        assert(type(ret) is type(s))
+        assert ret == b"defabc"
+
+    def test_char_repeat(self):
+        np_s = np.bytes_('abc')
+        np_u = np.str_('abc')
+        res_s = b'abc' * 5
+        res_u = 'abc' * 5
+        assert_(np_s * 5 == res_s)
+        assert_(np_u * 5 == res_u)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/core/tests/test_strings.py

@@ -0,0 +1,99 @@
+import pytest
+
+import operator
+import numpy as np
+
+from numpy.testing import assert_array_equal
+
+
+COMPARISONS = [
+    (operator.eq, np.equal, "=="),
+    (operator.ne, np.not_equal, "!="),
+    (operator.lt, np.less, "<"),
+    (operator.le, np.less_equal, "<="),
+    (operator.gt, np.greater, ">"),
+    (operator.ge, np.greater_equal, ">="),
+]
+
+
+@pytest.mark.parametrize(["op", "ufunc", "sym"], COMPARISONS)
+def test_mixed_string_comparison_ufuncs_fail(op, ufunc, sym):
+    arr_string = np.array(["a", "b"], dtype="S")
+    arr_unicode = np.array(["a", "c"], dtype="U")
+
+    with pytest.raises(TypeError, match="did not contain a loop"):
+        ufunc(arr_string, arr_unicode)
+
+    with pytest.raises(TypeError, match="did not contain a loop"):
+        ufunc(arr_unicode, arr_string)
+
+@pytest.mark.parametrize(["op", "ufunc", "sym"], COMPARISONS)
+def test_mixed_string_comparisons_ufuncs_with_cast(op, ufunc, sym):
+    arr_string = np.array(["a", "b"], dtype="S")
+    arr_unicode = np.array(["a", "c"], dtype="U")
+
+    # While there is no loop, manual casting is acceptable:
+    res1 = ufunc(arr_string, arr_unicode, signature="UU->?", casting="unsafe")
+    res2 = ufunc(arr_string, arr_unicode, signature="SS->?", casting="unsafe")
+
+    expected = op(arr_string.astype('U'), arr_unicode)
+    assert_array_equal(res1, expected)
+    assert_array_equal(res2, expected)
+
+
+@pytest.mark.parametrize(["op", "ufunc", "sym"], COMPARISONS)
+@pytest.mark.parametrize("dtypes", [
+        ("S2", "S2"), ("S2", "S10"),
+        ("<U1", "<U1"), ("<U1", ">U1"), (">U1", ">U1"),
+        ("<U1", "<U10"), ("<U1", ">U10")])
+@pytest.mark.parametrize("aligned", [True, False])
+def test_string_comparisons(op, ufunc, sym, dtypes, aligned):
+    # ensure native byte-order for the first view to stay within unicode range
+    native_dt = np.dtype(dtypes[0]).newbyteorder("=")
+    arr = np.arange(2**15).view(native_dt).astype(dtypes[0])
+    if not aligned:
+        # Make `arr` unaligned:
+        new = np.zeros(arr.nbytes + 1, dtype=np.uint8)[1:].view(dtypes[0])
+        new[...] = arr
+        arr = new
+
+    arr2 = arr.astype(dtypes[1], copy=True)
+    np.random.shuffle(arr2)
+    arr[0] = arr2[0]  # make sure one matches
+
+    expected = [op(d1, d2) for d1, d2 in zip(arr.tolist(), arr2.tolist())]
+    assert_array_equal(op(arr, arr2), expected)
+    assert_array_equal(ufunc(arr, arr2), expected)
+    assert_array_equal(np.compare_chararrays(arr, arr2, sym, False), expected)
+
+    expected = [op(d2, d1) for d1, d2 in zip(arr.tolist(), arr2.tolist())]
+    assert_array_equal(op(arr2, arr), expected)
+    assert_array_equal(ufunc(arr2, arr), expected)
+    assert_array_equal(np.compare_chararrays(arr2, arr, sym, False), expected)
+
+
+@pytest.mark.parametrize(["op", "ufunc", "sym"], COMPARISONS)
+@pytest.mark.parametrize("dtypes", [
+        ("S2", "S2"), ("S2", "S10"), ("<U1", "<U1"), ("<U1", ">U10")])
+def test_string_comparisons_empty(op, ufunc, sym, dtypes):
+    arr = np.empty((1, 0, 1, 5), dtype=dtypes[0])
+    arr2 = np.empty((100, 1, 0, 1), dtype=dtypes[1])
+
+    expected = np.empty(np.broadcast_shapes(arr.shape, arr2.shape), dtype=bool)
+    assert_array_equal(op(arr, arr2), expected)
+    assert_array_equal(ufunc(arr, arr2), expected)
+    assert_array_equal(np.compare_chararrays(arr, arr2, sym, False), expected)
+
+
+@pytest.mark.parametrize("str_dt", ["S", "U"])
+@pytest.mark.parametrize("float_dt", np.typecodes["AllFloat"])
+def test_float_to_string_cast(str_dt, float_dt):
+    float_dt = np.dtype(float_dt)
+    fi = np.finfo(float_dt)
+    arr = np.array([np.nan, np.inf, -np.inf, fi.max, fi.min], dtype=float_dt)
+    expected = ["nan", "inf", "-inf", repr(fi.max), repr(fi.min)]
+    if float_dt.kind == 'c':
+        expected = [f"({r}+0j)" for r in expected]
+
+    res = arr.astype(str_dt)
+    assert_array_equal(res, np.array(expected, dtype=str_dt))

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/core/tests/test_umath_accuracy.py

@@ -0,0 +1,75 @@
+import numpy as np
+import os
+from os import path
+import sys
+import pytest
+from ctypes import c_longlong, c_double, c_float, c_int, cast, pointer, POINTER
+from numpy.testing import assert_array_max_ulp
+from numpy.testing._private.utils import _glibc_older_than
+from numpy.core._multiarray_umath import __cpu_features__
+
+UNARY_UFUNCS = [obj for obj in np.core.umath.__dict__.values() if
+        isinstance(obj, np.ufunc)]
+UNARY_OBJECT_UFUNCS = [uf for uf in UNARY_UFUNCS if "O->O" in uf.types]
+UNARY_OBJECT_UFUNCS.remove(getattr(np, 'invert'))
+
+IS_AVX = __cpu_features__.get('AVX512F', False) or \
+        (__cpu_features__.get('FMA3', False) and __cpu_features__.get('AVX2', False))
+# only run on linux with AVX, also avoid old glibc (numpy/numpy#20448).
+runtest = (sys.platform.startswith('linux')
+           and IS_AVX and not _glibc_older_than("2.17"))
+platform_skip = pytest.mark.skipif(not runtest,
+                                   reason="avoid testing inconsistent platform "
+                                   "library implementations")
+
+# convert string to hex function taken from:
+# https://stackoverflow.com/questions/1592158/convert-hex-to-float #
+def convert(s, datatype="np.float32"):
+    i = int(s, 16)                   # convert from hex to a Python int
+    if (datatype == "np.float64"):
+        cp = pointer(c_longlong(i))           # make this into a c long long integer
+        fp = cast(cp, POINTER(c_double))  # cast the int pointer to a double pointer
+    else:
+        cp = pointer(c_int(i))           # make this into a c integer
+        fp = cast(cp, POINTER(c_float))  # cast the int pointer to a float pointer
+
+    return fp.contents.value         # dereference the pointer, get the float
+
+str_to_float = np.vectorize(convert)
+
+class TestAccuracy:
+    @platform_skip
+    def test_validate_transcendentals(self):
+        with np.errstate(all='ignore'):
+            data_dir = path.join(path.dirname(__file__), 'data')
+            files = os.listdir(data_dir)
+            files = list(filter(lambda f: f.endswith('.csv'), files))
+            for filename in files:
+                filepath = path.join(data_dir, filename)
+                with open(filepath) as fid:
+                    file_without_comments = (r for r in fid if not r[0] in ('$', '#'))
+                    data = np.genfromtxt(file_without_comments,
+                                         dtype=('|S39','|S39','|S39',int),
+                                         names=('type','input','output','ulperr'),
+                                         delimiter=',',
+                                         skip_header=1)
+                    npname = path.splitext(filename)[0].split('-')[3]
+                    npfunc = getattr(np, npname)
+                    for datatype in np.unique(data['type']):
+                        data_subset = data[data['type'] == datatype]
+                        inval  = np.array(str_to_float(data_subset['input'].astype(str), data_subset['type'].astype(str)), dtype=eval(datatype))
+                        outval = np.array(str_to_float(data_subset['output'].astype(str), data_subset['type'].astype(str)), dtype=eval(datatype))
+                        perm = np.random.permutation(len(inval))
+                        inval = inval[perm]
+                        outval = outval[perm]
+                        maxulperr = data_subset['ulperr'].max()
+                        assert_array_max_ulp(npfunc(inval), outval, maxulperr)
+
+    @pytest.mark.parametrize("ufunc", UNARY_OBJECT_UFUNCS)
+    def test_validate_fp16_transcendentals(self, ufunc):
+        with np.errstate(all='ignore'):
+            arr = np.arange(65536, dtype=np.int16)
+            datafp16 = np.frombuffer(arr.tobytes(), dtype=np.float16)
+            datafp32 = datafp16.astype(np.float32)
+            assert_array_max_ulp(ufunc(datafp16), ufunc(datafp32),
+                    maxulp=1, dtype=np.float16)

LTA_openwebtext_dualt/mini_owt_logdirichlet/logs/infer_lm1b_lr3e3_nobottleneck_step97k_decode1024_ema_selfcond_20260613_223847.log

@@ -0,0 +1,31 @@
+[start] 2026-06-13T22:38:47+00:00
+checkpoint=runs/lm1b_t5_pack_len128_C1_to_1024_pow1_d768_l12_h12_gbs512_4gpu_50ep_lr3e3_ema0p9999_elfopt_not5_nobottleneck_unfixed_norm_stateprobadd_selfcond_ce_fast_20260611_232614/step_097000.pt
+use_ema=1
+step=097000
+decode_steps=1024
+n=64 chunk_n=8 gpu=0
+out_base=/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/docs/lta_samples/metrics_20260614
+[2026-06-13T22:38:47+00:00] infer step=097000 decode=1024 -> /e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/docs/lta_samples/metrics_20260614/lm1b_len128_lr3e3_not5_nobottleneck_step97k_selfcond_step097000_ema_sc1p0_decode1024_n64
+[2026-06-13T22:38:47+00:00] run decode=1024 chunk=0 n=8 seed=123
+[2026-06-13T22:39:03+00:00] done decode=1024 chunk=0
+[2026-06-13T22:39:03+00:00] run decode=1024 chunk=1 n=8 seed=124
+[2026-06-13T22:39:19+00:00] done decode=1024 chunk=1
+[2026-06-13T22:39:19+00:00] run decode=1024 chunk=2 n=8 seed=125
+[2026-06-13T22:39:34+00:00] done decode=1024 chunk=2
+[2026-06-13T22:39:34+00:00] run decode=1024 chunk=3 n=8 seed=126
+[2026-06-13T22:39:50+00:00] done decode=1024 chunk=3
+[2026-06-13T22:39:50+00:00] run decode=1024 chunk=4 n=8 seed=127
+[2026-06-13T22:40:06+00:00] done decode=1024 chunk=4
+[2026-06-13T22:40:06+00:00] run decode=1024 chunk=5 n=8 seed=128
+[2026-06-13T22:40:21+00:00] done decode=1024 chunk=5
+[2026-06-13T22:40:21+00:00] run decode=1024 chunk=6 n=8 seed=129
+[2026-06-13T22:40:37+00:00] done decode=1024 chunk=6
+[2026-06-13T22:40:37+00:00] run decode=1024 chunk=7 n=8 seed=130
+[2026-06-13T22:40:53+00:00] done decode=1024 chunk=7
+merged 64 samples -> /e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/docs/lta_samples/metrics_20260614/lm1b_len128_lr3e3_not5_nobottleneck_step97k_selfcond_step097000_ema_sc1p0_decode1024_n64/sc1p0/samples64.txt
+loading scorer /e2e-data/evad-tech-vla/wanghan58/models/flowtext_scorers/gpt2-large-standard dtype=fp32 device=cuda
+run	kind	ppl	mean_entropy	distinct_1	distinct_2	top_token_mass	eos_rows	eos_total	ppl_tokens	t5_tokens	path
+sc1p0	raw_full	46.28551224661415	4.071705804822175	0.17464549751545266	0.5646060606060606	0.10374500060598715	64	128	6991	8251	/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/docs/lta_samples/metrics_20260614/lm1b_len128_lr3e3_not5_nobottleneck_step97k_selfcond_step097000_ema_sc1p0_decode1024_n64/sc1p0
+sc1p0	pre_eos	nan	0.0	0.015625	0.015873015873015872	1.0	0	0	0	64	/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/docs/lta_samples/metrics_20260614/lm1b_len128_lr3e3_not5_nobottleneck_step97k_selfcond_step097000_ema_sc1p0_decode1024_n64/sc1p0
+[2026-06-13T22:41:01+00:00] done
+[exit] 2026-06-13T22:41:01+00:00 rc=0

LTA_openwebtext_dualt/mini_owt_logdirichlet/logs/infer_owt_lr3e3_not5_nobottleneck_250k_decode64_ema_20260613_225804.log

@@ -0,0 +1,31 @@
+[start] 2026-06-13T22:58:04+00:00 tag=nobottleneck gpu=3
+checkpoint=runs/owt_t5_elftokenized_full_len1024_C1_to_1024_pow1_d768_l12_h12_gbs512_2x8gpu_50ep_lr3e3_ema0p9999_elfopt_not5_nobottleneck_unfixed_norm_stateprobadd_selfcond_ce_fast_20260611_220316/step_250000.pt
+use_ema=1
+step=250000
+decode_steps=64
+n=64 chunk_n=8 gpu=3
+out_base=/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/docs/lta_samples/metrics_20260614
+[2026-06-13T22:58:04+00:00] infer step=250000 decode=64 -> /e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/docs/lta_samples/metrics_20260614/owt_lr3e3_not5_nobottleneck_step250000_ema_sc1p0_decode64_n64
+[2026-06-13T22:58:04+00:00] run decode=64 chunk=0 n=8 seed=123
+[2026-06-13T22:58:15+00:00] done decode=64 chunk=0
+[2026-06-13T22:58:15+00:00] run decode=64 chunk=1 n=8 seed=124
+[2026-06-13T22:58:25+00:00] done decode=64 chunk=1
+[2026-06-13T22:58:25+00:00] run decode=64 chunk=2 n=8 seed=125
+[2026-06-13T22:58:36+00:00] done decode=64 chunk=2
+[2026-06-13T22:58:36+00:00] run decode=64 chunk=3 n=8 seed=126
+[2026-06-13T22:58:46+00:00] done decode=64 chunk=3
+[2026-06-13T22:58:46+00:00] run decode=64 chunk=4 n=8 seed=127
+[2026-06-13T22:58:56+00:00] done decode=64 chunk=4
+[2026-06-13T22:58:56+00:00] run decode=64 chunk=5 n=8 seed=128
+[2026-06-13T22:59:06+00:00] done decode=64 chunk=5
+[2026-06-13T22:59:06+00:00] run decode=64 chunk=6 n=8 seed=129
+[2026-06-13T22:59:16+00:00] done decode=64 chunk=6
+[2026-06-13T22:59:16+00:00] run decode=64 chunk=7 n=8 seed=130
+[2026-06-13T22:59:26+00:00] done decode=64 chunk=7
+merged 64 samples -> /e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/docs/lta_samples/metrics_20260614/owt_lr3e3_not5_nobottleneck_step250000_ema_sc1p0_decode64_n64/sc1p0/samples64.txt
+loading scorer /e2e-data/evad-tech-vla/wanghan58/models/flowtext_scorers/gpt2-large-standard dtype=fp32 device=cuda
+run	kind	ppl	mean_entropy	distinct_1	distinct_2	top_token_mass	eos_rows	eos_total	ppl_tokens	t5_tokens	path
+sc1p0	raw_full	14.022381162335854	4.914212015412217	0.049311139252094044	0.3446339751609655	0.03200952046747937	64	64	64840	65543	/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/docs/lta_samples/metrics_20260614/owt_lr3e3_not5_nobottleneck_step250000_ema_sc1p0_decode64_n64/sc1p0
+sc1p0	pre_eos	15.70514521711768	4.941422532573047	0.05089792060491494	0.35577119992438444	0.03304977945809704	0	0	60677	63480	/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/docs/lta_samples/metrics_20260614/owt_lr3e3_not5_nobottleneck_step250000_ema_sc1p0_decode64_n64/sc1p0
+[2026-06-13T22:59:40+00:00] done
+[exit] 2026-06-13T22:59:40+00:00 tag=nobottleneck rc=0