Changelog

0.19.0 (2026-05-15)

New feature

• The extensions module now publicly exposes the is_nullable_numeric_dtype() and get_data() functions.

0.18.2 (2026-05-06)

Bug fix

• Fixed an issue whereby any unused inputs dropped via the drop_unused argument to ndonnx.build() were still included in ndonnx_schema.

0.18.1 (2026-04-22)

Bug fix

• Fixed a regression that forbade using the modulo operator between integers and floats.

0.18.0 (2026-04-07)

New feature

• ndonnx.take() and ndonnx.take_along_axis() can now be used with datetime and timedelta data types.

Bug fix

• ndonnx.take() and ndonnx.take_along_axis() now validate their indices and axis arguments in adherence with the array-api standard.

0.17.4 (2026-03-23)

Bug fixes

• Reverse true division now works for nullable numeric arrays, so expressions such as 1 / arr behave correctly.

• ndonnx.result_type() now raises ValueError if called without any arguments.

• Invalid array-based slice bounds now raise IndexError during indexing.

• Datetime put operations now reject values with a mismatched dtype.

• Boolean mask __setitem__ now raises TypeError when the value has too many dimensions, matching NumPy.

0.17.3 (2026-03-10)

Bug fix

• Integer floordiv (//) is now computed correctly even for very large integers.

0.17.2 (2026-02-13)

Other change

• ndonnx.Array.dynamic_size() now produces more efficient graphs.

0.17.1 (2025-12-05)

Bug fix

• The masks of nullable arrays are now concatenated correctly.

0.17.0 (2025-12-01)

Breaking change

• ndonnx.clip() now behaves like NumPy when an integer dtype array and float type min or max values are given.

Other change

• Reduction operations now produce more efficient graphs.

0.16.0 (2025-08-20)

Bug fix

• ndonnx.repeat() now works correctly for zero-sized inputs on the minimum supported onnxruntime 1.20.1.

New feature

• ndonnx.max() (ndonnx.min()) now returns the minimum (maximum) value for the input data type if the reduction takes place over a zero-sized input.

• ndonnx.arange() now also allows arrays as inputs.

0.15.0 (2025-08-13)

New feature

• ndonnx.repeat() is now implemented for all built-in data types.

Other changes

• Improve stacklevel presented of various deprecation warnings.

0.14.0 (2025-07-21)

Bug fixes

• Fix a bug in ndonnx.Array.__setitem__() that occurred when all of the following applied: - An Ellipsis was part of the key - The Ellipsis expanded to at least one dimension - The Ellipsis was not the last element of the key - The assigned value was not a scalar or 1D array with length 1.

• ndonnx.Array.dynamic_shape() now returns a rank-0 array for all input ranks.

• The error message of the IndexingError raise by ndonnx.Array.__setitem__() when providing a tuple-key containing int64-arrays is now accurate.

• Using slice objects in the ndonnx.Array.__setitem__() no longer require value propagation.

• ndonnx.Array.__setitem__() now correctly handles boolean masks for arrays of two or more dimensions.

• Operations between NumPy and ndonnx arrays now correctly call the reverse dunder methods such as ndonnx.Array.__radd__() where appropriate.

• Operations between numpy.generic and ndonnx.Array objects now follow the regular type promotion logic.

• The following dunder operations on ndonnx.Array now correctly return NotImplemented if one of the operands is not of type numpy.ndarray | ndonnx.Array | bool | str | int | float: __add__, __radd__, __and__, __rand__, __floordiv__, __rfloordiv__, __ge__, __gt__, __le__, __lshift__, __rlshift__, __lt__, __matmul__, __rmatmul__, __mod__, __rmod__, __mul__, __rmul__, __or__, __ror__, __pow__, __rpow__, __rshift__, __rrshift__, __sub__, __rsub__, __truediv__, __rtruediv__, __xor__, __rxor__, __eq__, __ne__

New workarounds for missing onnxruntime implementations

• ndx.any() and ndx.all() now correctly propagate values even if the input is zero-sized.

• ndonnx.arange() now follows NumPy’s semantics for extremely large start, stop, and step values.

• ndonnx.min(), ndonnx.max(), ndonnx.minimum(), ndonnx.maximum(), and ndonnx.clip() now produce correct results for very large values in int64 arrays.

0.13.0 (2025-05-27)

Bug fixes

• ndonnx.concat() no longer raises an error if axis=None, the resulting data type is int32 or int64, and one of the provided arrays is zero-sized.

• ndonnx.__array_namespace_info__.capabilities() now reports the number of supported dimensions via the "max dimensions" entry rather than "max rank".

• Add missing onnxruntime workaround for uint32 inputs to ndonnx.min() and ndonnx.max().

• Fix array instantiation with ndonnx.asarray() and very large Python integers for uint64 data types.

• Fix passing an Python scalar as the second argument to ndonnx.where().

• Calling ndonnx.roll() on zero-sized inputs no longer causes a segfault on Linux.

New features

• The ndonnx.TimeDelta64DType and ndonnx.DateTime64DType gained support for milli and microseconds as units.

• ndonnx.where() now promotes time units between the two branches.

• Addition, multiplication, division, and subtraction between arrays with timedelta or datetime data types now support promotion between time units.

• Comparison operations between arrays with timedelta or datetime data types now support promotion between time units.

• Added ndonnx.__array_api_version__ reporting the latest supported version of the Array API specification.

0.12.0 (2025-05-15)

Bug fix

• The modulo operator (%) now correctly follows Python’s semantics if the second argument is negative.

New features

• Support the 2024.12 version of the Array API (except nextafter).

0.11.0 (2025-05-08)

Bug fix

• ndonnx.mean(), ndonnx.var(), and ndonnx.std() now produce correct results if axes outside the range of [-1, 1] are given.

Other change

• Address various deprecation warnings.

0.10.1 (2025-04-01)

Technical release to address a PyPI publishing issue.

0.10.0 (2025-04-01)

Breaking change

• Removed the deprecated ndonnx.promote_nullable() function. Use ndonnx.extensions.make_nullable() instead.

• The existing (private) tooling to define custom data types has been removed.

• The ndonnx.Array.len() member function has been removed since it is neither defined by numpy.ndarray nor the Array-API.

• ndonnx.Array.size() now returns int | None in accordance to with the Array-API rather than a Array instance.

Bug fixes

• The following functions now correctly take the keepdims argument into account: - ndonnx.Array.sum() - ndonnx.Array.prod() - ndonnx.Array.max() - ndonnx.Array.min() - ndonnx.Array.all() - ndonnx.Array.any()

New features

• ndonnx gained partial support for float16 data types.

• The following functions are now exposed in the ndonnx namespace in accordance to the Array-API standard: - ndonnx.__array_namespace_info__() - ndonnx.maximum() - ndonnx.minimum() - ndonnx.meshgrid() - ndonnx.moveaxis() - ndonnx.tile() - ndonnx.unstack() - ndonnx.vecdot()

• The newly added ndonnx.extensions module exposes the following functions: - datetime_to_year_month_day() - fill_null() - get_mask() - is_float_dtype() - is_integer_dtype() - is_nullable_dtype() - is_nullable_float_dtype() - is_nullable_integer_dtype() - is_numeric_dtype() - is_onnx_dtype() - is_signed_integer_dtype() - s_time_unit() - is_unsigned_integer_dtype() - isin() - make_nullable() - put() - shape() (deprecated in favor of ndonnx.Array.dynamic_shape()) - static_map()

• Arrays now expose the ndonnx.Array.device() property to improve Array API compatibility. Note that serializing an ONNX model inherently postpones device placement decisions to the runtime so currently one abstract device is supported.

• The Array object gained the following member functions: - disassemble() - dynamic_shape() - dynamic_size() - unwrap_numpy() - unwrap_spox()

Deprecations

• ndonnx.array() is deprecated in favor of ndonnx.argument().

• ndonnx.additional is deprecated in favor of ndonnx.extensions().

• ndonnx.from_spox_var() is deprecated in favor of ndonnx.asarray().

• ndonnx.Nullable is deprecated in favor of ndonnx.extensions.is_nullable_dtype()

• ndonnx.NullableFloating is deprecated in favor of ndonnx.extensions.is_nullable_float_dtype()

• ndonnx.NullableIntegral is deprecated in favor of ndonnx.extensions.is_nullable_integer_dtype()

• ndonnx.Floating is deprecated in favor of ndonnx.extensions.is_float_dtype()

• ndonnx.Integral is deprecated in favor of ndonnx.extensions.is_integer_dtype()

• ndonnx.Numerical is deprecated in favor of ndonnx.extensions.is_numeric_dtype()

• ndonnx.CoreType is deprecated in favor of ndonnx.extensions.is_onnx_dtype()

• ndonnx.NullableCore is deprecated in favor of ndonnx.extensions.is_nullable_dtype()

• ndonnx.UnsupportedOperationError is deprecated in favor of TypeError

• ndonnx.CastError is deprecated in favor of TypeError

Other changes

• make_nullable() can no longer be used for custom data types.

0.9.3 (2024-10-25)

• Reduced the number of unnecessary casts in ndonnx.argmax() and ndonnx.argmin().

0.9.2 (2024-10-03)

• Technical release with source distribution.

0.9.1 (2024-10-01)

Bug fix

• Fixed a bug in the construction of nullable arrays using ndonnx.asarray() where the shape of the null field would not match the values field if the provided np.ma.MaskedArray’s mask was scalar.

• Fixed a bug in the implementation of ndonnx.ones_like() where the static shape was being used to construct the array of ones.

0.9.0 (2024-08-30)

New features

• User defined data types can now define how arrays with that dtype are constructed by implementing the make_array function.

• User defined data types can now define how they are indexed (via __getitem__) by implementing the getitem function.

• ndonnx.NullableCore is now public, encapsulating nullable variants of `CoreType`s exported by ndonnx.

Bug fixes

• Various operations that depend on the array’s shape have been updated to work correctly with lazy arrays.

• ndonnx.cumulative_sum() now correctly applies the include_initial parameter and works around missing onnxruntime kernels for unsigned integral types.

• ndonnx.additional.make_nullable() applies broadcasting to the provided null array (instead of reshape like it did previously). This allows writing make_nullable(x, False) to turn an array into nullable.

• User-defined data types that implement ndonnx._core.UniformShapeOperations may now implement ndonnx.where() without requiring both data types be promotable.

Breaking change

• Iterating over dynamic dimensions of Array is no longer allowed since it commonly lead to infinite loops when used without an explicit break condition.

0.8.0 (2024-08-22)

Bug fixes

• Fixes parsing numpy arrays of type object (consisting of strings) as utf8. Previously this worked correctly only for 1d arrays.

Breaking change

• ndonnx.Array.shape() now strictly returns a tuple[int | None, ...], with unknown dimensions denoted by None. This relies on ONNX shape inference for lazy arrays.

0.7.0 (2024-08-12)

New features

• Expose the ndonnx.isdtype() function.

• Custom data types can now override array functions:

  • ndonnx.zeros()

  • ndonnx.zeros_like()

  • ndonnx.ones()

  • ndonnx.ones_like()

  • ndonnx.full()

  • ndonnx.full_like()

  • ndonnx.arange()

  • ndonnx.arange()

  • ndonnx.eye()

  • ndonnx.tril()

  • ndonnx.triu()

  • ndonnx.linspace()

  • ndonnx.where()

• The ndonnx._experimental.UniformShapeOperations now provides implementations of shape operations that are generic across all data types where each constituent field has the same shape (that of the overall array).

Other changes

• Fixed various deprecation warnings.

• Invoking a function using arrays with data types that lack a corresponding implementation now raise a UnsupportedOperationError.

Bug fixes

• Numerical operations like sin() now raise UnsupportedOperationError when invoked using invalid data types like ndx.utf8 rather than implicitly casting.

• Fixes bug causing a promotion error when implementing numerical operations like add() that involve type promotion.

• Fixes scalar promotion logic to more accurately reflect the Array API standard. Promotion requires at least one array to be present and scalars adopt the dtype of the arrays being promoted with it. ndx.utf8 and ndx.nutf8 cannot be promoted with any other dtypes.

• Fixes failure when broadcasting nullable data type arrays together in broadcast_arrays().

0.6.1 (2024-07-12)

Bug fixes

• Division now complies more strictly with the Array API standard by returning a floating-point result regardless of input data types.

0.6.0 (2024-07-11)

Other changes

• ndonnx.promote_nullable is now publicly exported.

0.5.0 (2024-07-01)

Other changes

• ndonnx now exports type annotations.

Bug fixes

• __array_namespace__ now accepts the optional api_version argument to specify the version of the Array API to use.

0.4.0 (2024-05-16)

Breaking changes

• The constant propagated value is no longer accessed from the eager_value property but instead the to_numpy() method.

• Non Array API functions have been moved to the ndonnx.additional namespace.