doc/source/user/basics.interoperability.rst - main - niceSOFT/numpy

doc/source/user/basics.interoperability.rst
547 строк21 KB
majianhan
DOC: clarify DLPack GPU tensor example with device='cpu' workaround
11 мар 2026, 15:31
11 мар 2026, 15:31f603c46

.. _basics.interoperability:

***************************
Interoperability with NumPy
***************************

NumPy's ndarray objects provide both a high-level API for operations on
array-structured data and a concrete implementation of the API based on
:ref:`strided in-RAM storage <arrays>`. While this API is powerful and fairly
general, its concrete implementation has limitations. As datasets grow and NumPy
becomes used in a variety of new environments and architectures, there are cases
where the strided in-RAM storage strategy is inappropriate, which has caused
different libraries to reimplement this API for their own uses. This includes
GPU arrays (CuPy_), Sparse arrays (`scipy.sparse`, `PyData/Sparse <Sparse_>`_)
and parallel arrays (Dask_ arrays) as well as various NumPy-like implementations
in deep learning frameworks, like TensorFlow_ and PyTorch_. Similarly, there are
many projects that build on top of the NumPy API for labeled and indexed arrays
(XArray_), automatic differentiation (JAX_), masked arrays (`numpy.ma`),
physical units (astropy.units_, pint_, unyt_), among others that add additional
functionality on top of the NumPy API.

Yet, users still want to work with these arrays using the familiar NumPy API and
reuse existing code with minimal (ideally zero) porting overhead. With this
goal in mind, various protocols are defined for implementations of
multi-dimensional arrays with high-level APIs matching NumPy.

Broadly speaking, there are three groups of features used for interoperability
with NumPy:

1. Methods of turning a foreign object into an ndarray;
2. Methods of deferring execution from a NumPy function to another array
   library;
3. Methods that use NumPy functions and return an instance of a foreign object.

We describe these features below.


1. Using arbitrary objects in NumPy
-----------------------------------

The first set of interoperability features from the NumPy API allows foreign
objects to be treated as NumPy arrays whenever possible. When NumPy functions
encounter a foreign object, they will try (in order):

1. The buffer protocol, described :py:doc:`in the Python C-API documentation
   <python:c-api/buffer>`.
2. The ``__array_interface__`` protocol, described
   :ref:`in this page <arrays.interface>`. A precursor to Python's buffer
   protocol, it defines a way to access the contents of a NumPy array from other
   C extensions.
3. The ``__array__()`` method, which asks an arbitrary object to convert
   itself into an array.

For both the buffer and the ``__array_interface__`` protocols, the object
describes its memory layout and NumPy does everything else (zero-copy if
possible). If that's not possible, the object itself is responsible for
returning a ``ndarray`` from ``__array__()``.

:doc:`DLPack <dlpack:index>` is yet another protocol to convert foreign objects
to NumPy arrays in a language and device agnostic manner. NumPy doesn't implicitly
convert objects to ndarrays using DLPack. It provides the function
`numpy.from_dlpack` that accepts any object implementing the ``__dlpack__`` method
and outputs a NumPy ndarray (which is generally a view of the input object's data
buffer). The :ref:`dlpack:python-spec` page explains the ``__dlpack__`` protocol
in detail.

``dtype`` interoperability
~~~~~~~~~~~~~~~~~~~~~~~~~~
Similar to ``__array__()`` for array objects, defining ``__numpy_dtype__``
allows a custom dtype object to be interoperable with NumPy.
The ``__numpy_dtype__`` must return a NumPy dtype instance (note that
``np.float64`` is not a dtype instance, ``np.dtype(np.float64)`` is).

.. versionadded:: 2.4
   Before NumPy 2.4 a ``.dtype`` attribute was treated similarly. As of NumPy 2.4
   both is accepted and implementing ``__numpy_dtype__`` prevents ``.dtype``
   from being checked.

The array interface protocol
~~~~~~~~~~~~~~~~~~~~~~~~~~~~

The :ref:`array interface protocol <arrays.interface>` defines a way for
array-like objects to reuse each other's data buffers. Its implementation
relies on the existence of the following attributes or methods:

-  ``__array_interface__``: a Python dictionary containing the shape, the
   element type, and optionally, the data buffer address and the strides of an
   array-like object;
-  ``__array__()``: a method returning the NumPy ndarray copy or a view of an
   array-like object;

The ``__array_interface__`` attribute can be inspected directly:

 >>> import numpy as np
 >>> x = np.array([1, 2, 5.0, 8])
 >>> x.__array_interface__
 {'data': (94708397920832, False), 'strides': None, 'descr': [('', '<f8')], 'typestr': '<f8', 'shape': (4,), 'version': 3}

The ``__array_interface__`` attribute can also be used to manipulate the object
data in place:

 >>> class wrapper():
 ...     pass
 ...
 >>> arr = np.array([1, 2, 3, 4])
 >>> buf = arr.__array_interface__
 >>> buf
 {'data': (140497590272032, False), 'strides': None, 'descr': [('', '<i8')], 'typestr': '<i8', 'shape': (4,), 'version': 3}
 >>> buf['shape'] = (2, 2)
 >>> w = wrapper()
 >>> w.__array_interface__ = buf
 >>> new_arr = np.array(w, copy=False)
 >>> new_arr
 array([[1, 2],
        [3, 4]])

We can check that ``arr`` and ``new_arr`` share the same data buffer:

 >>> new_arr[0, 0] = 1000
 >>> new_arr
 array([[1000,    2],
        [   3,    4]])
 >>> arr
 array([1000, 2, 3, 4])


.. _dunder_array.interface:

The ``__array__()`` method
~~~~~~~~~~~~~~~~~~~~~~~~~~

The `__array__() <../reference/arrays.classes.html#numpy.class.\_\_array\_\_>`__ method ensures that any NumPy-like object (an array, any
object exposing the array interface, an object whose ``__array__()`` method
returns an array or any nested sequence) that implements it can be used as a
NumPy array. If possible, this will mean using ``__array__()`` to create a NumPy
ndarray view of the array-like object. Otherwise, this copies the data into a
new ndarray object. This is not optimal, as coercing arrays into ndarrays may
cause performance problems or create the need for copies and loss of metadata,
as the original object and any attributes/behavior it may have had, is lost.

The signature of the method should be ``__array__(self, dtype=None, copy=None)``.
If a passed ``dtype`` isn't ``None`` and different than the object's data type,
a casting should happen to a specified type. If ``copy`` is ``None``, a copy
should be made only if ``dtype`` argument enforces it. For ``copy=True``, a copy
should always be made, where ``copy=False`` should raise an exception if a copy
is needed.

If a class implements the old signature ``__array__(self)``, for ``np.array(a)``
a warning will be raised saying that ``dtype`` and ``copy`` arguments are missing.

The DLPack Protocol
~~~~~~~~~~~~~~~~~~~

The :doc:`DLPack <dlpack:index>` protocol defines a memory-layout of
strided n-dimensional array objects. It offers the following syntax
for data exchange:

1. A `numpy.from_dlpack` function, which accepts (array) objects with a
   ``__dlpack__`` method and uses that method to construct a new array
   containing the data from ``x``.
2. ``__dlpack__(self, stream=None)`` and ``__dlpack_device__`` methods on the
   array object, which will be called from within ``from_dlpack``, to query
   what device the array is on (may be needed to pass in the correct
   stream, e.g. in the case of multiple GPUs) and to access the data.

Unlike the buffer protocol, DLPack allows exchanging arrays containing data on
devices other than the CPU (e.g. Vulkan or GPU). Since NumPy only supports CPU,
it can only convert objects whose data exists on the CPU. But other libraries,
like PyTorch_ and CuPy_, may exchange data on GPU using this protocol.


2. Operating on foreign objects without converting
--------------------------------------------------

A second set of methods defined by the NumPy API allows us to defer the
execution from a NumPy function to another array library.

Consider the following function.

 >>> import numpy as np
 >>> def f(x):
 ...     return np.mean(np.exp(x))

Note that `np.exp <numpy.exp>` is a :ref:`ufunc <ufuncs-basics>`, which means
that it operates on ndarrays in an element-by-element fashion. On the other
hand, `np.mean <numpy.mean>` operates along one of the array's axes.

We can apply ``f`` to a NumPy ndarray object directly:

 >>> x = np.array([1, 2, 3, 4])
 >>> f(x)
 21.1977562209304

We would like this function to work equally well with any NumPy-like array
object.

NumPy allows a class to indicate that it would like to handle computations in a
custom-defined way through the following interfaces:

-  ``__array_ufunc__``: allows third-party objects to support and override
   :ref:`ufuncs <ufuncs-basics>`.
-  ``__array_function__``: a catch-all for NumPy functionality that is not
   covered by the ``__array_ufunc__`` protocol for universal functions.

As long as foreign objects implement the ``__array_ufunc__`` or
``__array_function__`` protocols, it is possible to operate on them without the
need for explicit conversion.

The ``__array_ufunc__`` protocol
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

A :ref:`universal function (or ufunc for short) <ufuncs-basics>` is a
“vectorized” wrapper for a function that takes a fixed number of specific inputs
and produces a fixed number of specific outputs. The output of the ufunc (and
its methods) is not necessarily an ndarray, if not all input arguments are
ndarrays. Indeed, if any input defines an ``__array_ufunc__`` method, control
will be passed completely to that function, i.e., the ufunc is overridden. The
``__array_ufunc__`` method defined on that (non-ndarray) object has access to
the NumPy ufunc. Because ufuncs have a well-defined structure, the foreign
``__array_ufunc__`` method may rely on ufunc attributes like ``.at()``,
``.reduce()``, and others.

A subclass can override what happens when executing NumPy ufuncs on it by
overriding the default ``ndarray.__array_ufunc__`` method. This method is
executed instead of the ufunc and should return either the result of the
operation, or ``NotImplemented`` if the operation requested is not implemented.

The ``__array_function__`` protocol
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

To achieve enough coverage of the NumPy API to support downstream projects,
there is a need to go beyond ``__array_ufunc__`` and implement a protocol that
allows arguments of a NumPy function to take control and divert execution to
another function (for example, a GPU or parallel implementation) in a way that
is safe and consistent across projects.

The semantics of ``__array_function__`` are very similar to ``__array_ufunc__``,
except the operation is specified by an arbitrary callable object rather than a
ufunc instance and method. For more details, see :ref:`NEP18`.


3. Returning foreign objects
----------------------------

A third type of feature set is meant to use the NumPy function implementation
and then convert the return value back into an instance of the foreign object.
The ``__array_finalize__`` and ``__array_wrap__`` methods act behind the scenes
to ensure that the return type of a NumPy function can be specified as needed.

The ``__array_finalize__`` method is the mechanism that NumPy provides to allow
subclasses to handle the various ways that new instances get created. This
method is called whenever the system internally allocates a new array from an
object which is a subclass (subtype) of the ndarray. It can be used to change
attributes after construction, or to update meta-information from the “parent.”

The ``__array_wrap__`` method “wraps up the action” in the sense of allowing any
object (such as user-defined functions) to set the type of its return value and
update attributes and metadata. This can be seen as the opposite of the
``__array__`` method. At the end of every object that implements
``__array_wrap__``, this method is called on the input object with the highest
*array priority*, or the output object if one was specified. The
``__array_priority__`` attribute is used to determine what type of object to
return in situations where there is more than one possibility for the Python
type of the returned object. For example, subclasses may opt to use this method
to transform the output array into an instance of the subclass and update
metadata before returning the array to the user.

For more information on these methods, see :ref:`basics.subclassing` and
:ref:`specific-array-subtyping`.


Interoperability examples
-------------------------

Example: Pandas ``Series`` objects
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Consider the following:

 >>> import pandas as pd
 >>> ser = pd.Series([1, 2, 3, 4])
 >>> type(ser)
 pandas.core.series.Series

Now, ``ser`` is **not** an ndarray, but because it
`implements the __array_ufunc__ protocol
<https://pandas.pydata.org/docs/user_guide/dsintro.html#dataframe-interoperability-with-numpy-functions>`__,
we can apply ufuncs to it as if it were an ndarray:

 >>> np.exp(ser)
    0     2.718282
    1     7.389056
    2    20.085537
    3    54.598150
    dtype: float64
 >>> np.sin(ser)
    0    0.841471
    1    0.909297
    2    0.141120
    3   -0.756802
    dtype: float64

We can even do operations with other ndarrays:

 >>> np.add(ser, np.array([5, 6, 7, 8]))
    0     6
    1     8
    2    10
    3    12
    dtype: int64
 >>> f(ser)
 21.1977562209304
 >>> result = ser.__array__()
 >>> type(result)
 numpy.ndarray


Example: PyTorch tensors
~~~~~~~~~~~~~~~~~~~~~~~~

`PyTorch <https://pytorch.org/>`__ is an optimized tensor library for deep
learning using GPUs and CPUs. PyTorch arrays are commonly called *tensors*.
Tensors are similar to NumPy's ndarrays, except that tensors can run on GPUs or
other hardware accelerators. In fact, tensors and NumPy arrays can often share
the same underlying memory, eliminating the need to copy data.

 >>> import torch
 >>> data = [[1, 2],[3, 4]]
 >>> x_np = np.array(data)
 >>> x_tensor = torch.tensor(data)

Note that ``x_np`` and ``x_tensor`` are different kinds of objects:

 >>> x_np
 array([[1, 2],
        [3, 4]])
 >>> x_tensor
 tensor([[1, 2],
         [3, 4]])

However, we can treat PyTorch tensors as NumPy arrays without the need for
explicit conversion:

 >>> np.exp(x_tensor)
 tensor([[ 2.7183,  7.3891],
         [20.0855, 54.5982]], dtype=torch.float64)

Also, note that the return type of this function is compatible with the initial
data type.

.. admonition:: Warning

   While this mixing of ndarrays and tensors may be convenient, it is not
   recommended. It will not work for non-CPU tensors, and will have unexpected
   behavior in corner cases. Users should prefer explicitly converting the
   ndarray to a tensor.

.. note::

   PyTorch does not implement ``__array_function__`` or ``__array_ufunc__``.
   Under the hood, the ``Tensor.__array__()`` method returns a NumPy ndarray as
   a view of the tensor data buffer. See `this issue
   <https://github.com/pytorch/pytorch/issues/24015>`__ and the
   `__torch_function__ implementation
   <https://github.com/pytorch/pytorch/blob/master/torch/overrides.py>`__
   for details.

Note also that we can see ``__array_wrap__`` in action here, even though
``torch.Tensor`` is not a subclass of ndarray::

   >>> import torch
   >>> t = torch.arange(4)
   >>> np.abs(t)
   tensor([0, 1, 2, 3])

PyTorch implements ``__array_wrap__`` to be able to get tensors back from NumPy
functions, and we can modify it directly to control which type of objects are
returned from these functions.

Example: CuPy arrays
~~~~~~~~~~~~~~~~~~~~

CuPy is a NumPy/SciPy-compatible array library for GPU-accelerated computing
with Python. CuPy implements a subset of the NumPy interface by implementing
``cupy.ndarray``, `a counterpart to NumPy ndarrays
<https://docs.cupy.dev/en/stable/reference/ndarray.html>`__.

 >>> import cupy as cp
 >>> x_gpu = cp.array([1, 2, 3, 4])

The ``cupy.ndarray`` object implements the ``__array_ufunc__`` interface. This
enables NumPy ufuncs to be applied to CuPy arrays (this will defer operation to
the matching CuPy CUDA/ROCm implementation of the ufunc):

 >>> np.mean(np.exp(x_gpu))
 array(21.19775622)

Note that the return type of these operations is still consistent with the
initial type:

 >>> arr = cp.random.randn(1, 2, 3, 4).astype(cp.float32)
 >>> result = np.sum(arr)
 >>> print(type(result))
 <class 'cupy._core.core.ndarray'>

See `this page in the CuPy documentation for details
<https://docs.cupy.dev/en/stable/reference/ufunc.html>`__.

``cupy.ndarray`` also implements the ``__array_function__`` interface, meaning
it is possible to do operations such as

 >>> a = np.random.randn(100, 100)
 >>> a_gpu = cp.asarray(a)
 >>> qr_gpu = np.linalg.qr(a_gpu)

CuPy implements many NumPy functions on ``cupy.ndarray`` objects, but not all.
See `the CuPy documentation
<https://docs.cupy.dev/en/stable/user_guide/difference.html>`__
for details.

Example: Dask arrays
~~~~~~~~~~~~~~~~~~~~

Dask is a flexible library for parallel computing in Python. Dask Array
implements a subset of the NumPy ndarray interface using blocked algorithms,
cutting up the large array into many small arrays. This allows computations on
larger-than-memory arrays using multiple cores.

Dask supports ``__array__()`` and ``__array_ufunc__``.

 >>> import dask.array as da
 >>> x = da.random.normal(1, 0.1, size=(20, 20), chunks=(10, 10))
 >>> np.mean(np.exp(x))
 dask.array<mean_agg-aggregate, shape=(), dtype=float64, chunksize=(), chunktype=numpy.ndarray>
 >>> np.mean(np.exp(x)).compute()
 5.090097550553843

.. note::

   Dask is lazily evaluated, and the result from a computation isn't computed
   until you ask for it by invoking ``compute()``.

See `the Dask array documentation
<https://docs.dask.org/en/stable/array.html>`__
and the `scope of Dask arrays interoperability with NumPy arrays
<https://docs.dask.org/en/stable/array.html#scope>`__ for details.

Example: DLPack
~~~~~~~~~~~~~~~

Several Python data science libraries implement the ``__dlpack__`` protocol.
Among them are PyTorch_ and CuPy_. A full list of libraries that implement
this protocol can be found on
:doc:`this page of DLPack documentation <dlpack:index>`.

Convert a PyTorch CPU tensor to NumPy array:

 >>> import torch
 >>> x_torch = torch.arange(5)
 >>> x_torch
 tensor([0, 1, 2, 3, 4])
 >>> x_np = np.from_dlpack(x_torch)
 >>> x_np
 array([0, 1, 2, 3, 4])
 >>> # note that x_np is a view of x_torch
 >>> x_torch[1] = 100
 >>> x_torch
 tensor([  0, 100,   2,   3,   4])
 >>> x_np
 array([  0, 100,   2,   3,   4])

The imported arrays are read-only so writing or operating in-place will fail:

 >>> x_np.flags.writeable
 False
 >>> x_np[1] = 1
 Traceback (most recent call last):
   File "<stdin>", line 1, in <module>
 ValueError: assignment destination is read-only

A copy must be created in order to operate on the imported arrays in-place, but
will mean duplicating the memory. Do not do this for very large arrays:

 >>> x_np_copy = x_np.copy()
 >>> x_np_copy.sort()  # works

.. note::

  GPU tensors cannot be directly zero-copy converted to NumPy arrays since
  NumPy does not support GPU devices. However, since DLPack v1, cross-device
  copy is supported via the ``device`` parameter:

   >>> x_torch = torch.arange(5, device='cuda')
   >>> np.from_dlpack(x_torch)  # fails: implicit device=None means same device
   Traceback (most recent call last):
     File "<stdin>", line 1, in <module>
   RuntimeError: Unsupported device in DLTensor.
   >>> np.from_dlpack(x_torch, device='cpu')  # works: explicit copy to CPU
   array([0, 1, 2, 3, 4])

  If both libraries support the device the data buffer is on, it is
  possible to use the ``__dlpack__`` protocol (e.g. PyTorch_ and CuPy_):

   >>> x_torch = torch.arange(5, device='cuda')
   >>> x_cupy = cupy.from_dlpack(x_torch)

Similarly, a NumPy array can be converted to a PyTorch tensor:

 >>> x_np = np.arange(5)
 >>> x_torch = torch.from_dlpack(x_np)

Read-only arrays cannot be exported:

 >>> x_np = np.arange(5)
 >>> x_np.flags.writeable = False
 >>> torch.from_dlpack(x_np)  # doctest: +ELLIPSIS
 Traceback (most recent call last):
   File "<stdin>", line 1, in <module>
   File ".../site-packages/torch/utils/dlpack.py", line 63, in from_dlpack
     dlpack = ext_tensor.__dlpack__()
 TypeError: NumPy currently only supports dlpack for writeable arrays

Further reading
---------------

-  :ref:`arrays.interface`
-  :ref:`basics.dispatch`
-  :ref:`special-attributes-and-methods` (details on the ``__array_ufunc__`` and
   ``__array_function__`` protocols)
-  :ref:`basics.subclassing` (details on the ``__array_wrap__`` and
   ``__array_finalize__`` methods)
-  :ref:`specific-array-subtyping` (more details on the implementation of
   ``__array_finalize__``, ``__array_wrap__`` and ``__array_priority__``)
-  :doc:`NumPy roadmap: interoperability <neps:roadmap>`
-  `PyTorch documentation on the Bridge with NumPy
   <https://pytorch.org/tutorials/beginner/blitz/tensor_tutorial.html#bridge-to-np-label>`__

.. _CuPy: https://cupy.dev/
.. _Sparse: https://sparse.pydata.org/
.. _Dask: https://docs.dask.org/
.. _TensorFlow: https://www.tensorflow.org/
.. _PyTorch: https://pytorch.org/
.. _XArray: https://xarray.dev/
.. _JAX: https://jax.readthedocs.io/
.. _astropy.units: https://docs.astropy.org/en/stable/units/
.. _pint: https://pint.readthedocs.io/
.. _unyt: https://unyt.readthedocs.io/