Array's dtype asarray(np_float32_array, device=Device("device1"))

Another change between 2.5 and 2.6. In 2.6 the dtype of an array can be changed when you array_api_strict.asarray(x, device=Device("device1"))

With v2.5 this snippet prints array_api_strict.float32, with v2.6 you get float64.

import numpy as np
import array_api_strict as xp

x = np.ones((2,3), dtype=np.float32)
y = xp.asarray(x, device=xp.Device("device1"))

print(y.dtype)

I think preserving the dtype (if it is supported on the target device) is what should happen (aka behaviour of v2.5). I think the default dtype is something to look at for functions like ones() and friends. For those there is no hint from the user what the type should be.

A fix could be to use:

    # numpy default dtype may differ; if so, adjust the dtype.
    # Only do this for inputs whose dtype was inferred by NumPy (Python
    # scalars / sequences). Inputs that carry their own dtype (arrays and
    # buffer-protocol objects) must have that dtype preserved.
    if dtype is None and device is not None and not _supports_buffer_protocol(obj):
        res_dtype = DType(res.dtype)

A fix could be to use:

Or perhaps more simply — mimicking the existing:

    if isinstance(obj, Array) and device is None:
        device = obj.device

we could have

    if isinstance(obj, np.ndarray) and dtype is None:
        _np_dtype = obj.dtype

I think we need _supports_buffer_protocol(obj) to catch other inputs that have a dtype attached, like numpy scalars or torch arrays (on the CPU).

Grr. _supports_buffer_protocol won't suffice, sadly, because torch does not support it, even on CPU.

In [1]: import array_api_strict as xp

In [2]: xp.__version__
Out[2]: '2.5'

In [3]: import torch

In [4]: xp._creation_functions._supports_buffer_protocol(torch.ones(3))    # oops
Out[4]: False

In [5]: xp.asarray(torch.ones(3))
Out[5]: Array([1., 1., 1.], dtype=array_api_strict.float32)

In [6]: torch.ones(3).device
Out[6]: device(type='cpu')

Special-casing np.ndarray won't do either as a bugfix because In[5] above: xp.asarray(torch.ones(3)) succeeds in 2.5.
ATM the only way around I see is to branch on hasattr(obj, "dtype"). Any better ideas?

Unless of course, we declare the 2.5 behavior a bug and declare that array-api-strict does not really support xp.asarray from alternative array libraries?

Unless of course, we declare the 2.5 behavior a bug and declare that array-api-strict does not really support xp.asarray from alternative array libraries?

If it is indeed the case that PyTorch does not support the buffer protocol (is memoryview(obj) the correct test for that?), then yes, it seems better for array-api-strict to not support whatever makes np.asarray work (that is presumably absent from the spec).

Here's what I believe is a backwards compatible fix: #223
Please do review / test / poke holes in it!

If it is indeed the case that PyTorch does not support the buffer protocol (is memoryview(obj) the correct test for that?), then yes, it seems better for array-api-strict to not support whatever makes np.asarray work (that is presumably absent from the spec).

I believe that yes and yes (pytorch does not support buffer protocol, and memoryview is the appropriate check for it).

That said, let's maybe consider making array-api-strict stricter in a regular release (and consider this together with __array__). Meanwhile I think gating the dtype adjustment dance on obj.dtype is makes it backwards compatible: we first call np.array(obj), that either succeeded or failed. If it failed, there's no change in 2.6 or here; if it succeeded, then two options: obj is a python object (then we adjust the dtype) or it's an array (and then numpy did the work already and res.dtype matches obj.dtype).

MAINT: adapt to array-api-strict 2.6 per-device dtype semantics

ENH: improve device handling more, add pytorch meta device testing (#…

Navigation Menu

Repository navigation

Array's dtype `asarray(np_float32_array, device=Device("device1"))` #222

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Array's dtype asarray(np_float32_array, device=Device("device1")) #222

Description

Activity

lucascolley commented on Jul 10, 2026

betatim commented on Jul 10, 2026

ev-br commented on Jul 10, 2026

ev-br commented on Jul 10, 2026

lucascolley commented on Jul 10, 2026

ev-br commented on Jul 10, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

Array's dtype `asarray(np_float32_array, device=Device("device1"))` #222