[Bug] NumPy and bfloat16

[jaketae]

While examining why test cases were failing for

pytest ivy_tests/test_ivy/test_frontends/test_torch/test_creation_ops.py -k 'test_torch_full'

I realized that the to_numpy() function has a special branch for handling bfloat16:

https://github.com/unifyai/ivy/blob/e4f6cb774f18b8789c99a1ac4df22f342e858764/ivy/functional/backends/torch/general.py#L86-L92

The line I'm struggling to parse is

return x.detach().cpu().numpy().astype("bfloat16")

Currently, NumPy has no support for bfloat16. This can easily be checked in the interpreter.

>>> import numpy as np
>>> x = np.random.randn(1)
>>> x.astype("float16")
array([-0.343], dtype=float16)
>>> x.astype("bfloat16")
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: data type 'bfloat16' not understood

There is an ongoing discussion in the NumPy repo regarding bfloat16 support. Currently, they are conservative with adopting bfloat16 and are waiting for an external proof-of-concept implementation from the community.

In short, why are we calling np.ndarray.astype("bfloat16") in the to_numpy() conversion? Thank you!

---

Edit: Another implicit failure comes from array_helpers.py, specifically the array_values() function. The function attempts to return a NumPy array, but when dtype is bfloat16, the return fails:

https://github.com/unifyai/ivy/blob/e4f6cb774f18b8789c99a1ac4df22f342e858764/ivy_tests/test_ivy/helpers/hypothesis_helpers/array_helpers.py#L775

This behavior is observed, for instance, when running

pytest ivy_tests/test_ivy/test_frontends/test_torch/test_tensor.py -k 'test_torch_tensor_property_ivy_array'

[jaketae]

I found some PRs that appeared relevant to this topic.

• Adding specific behaviour to to_numpy for bfloat16 #6362
• Revert "Adding specific behaviour to to_numpy for bfloat16" #6364
• Fixing to_numpy bfloat16 behaviour #6366
• Fixed to_numpy issue #6476

@aditya-pola @vedpatwardhan, do you have any insight you could share? Thank you!