[Keras 3 OpenVINO backend]: Support numpy.view operation

[goyaladitya05]

Request Description

This issue tracks adding support for numpy.view in the Keras OpenVINO backend.
The operation is currently not implemented in keras/backend/openvino, limiting NumPy parity for this backend.

I will submit a PR implementing this and enable the corresponding tests.

Ref: #34017

Feature Use Case

No response

Issue submission checklist

• The feature request or improvement must be related to OpenVINO

[rishi-sangare]

.take

I've reviewed the current stub and reference implementations in TF/JAX/PyTorch backends. Since OpenVINO's opset doesn't have a native bitcast operation, I plan to implement view by reinterpreting bytes through a uint8 intermediate — handling the three cases (same size, shrink, expand) similar to the TF backend's approach.

Will also re-enable the 4 excluded test_view tests and submit a PR to keras-team/keras.

[rishi-sangare]

.take

[goyaladitya05]

@rishi-sangare Are you working on this?

[rishi-sangare]

Hi @goyaladitya05, yes I'm actively working on this. Here's my implementation plan:

Since OpenVINO's opset doesn't have a native bitcast, I'm implementing view by reinterpreting bytes through a uint8 intermediate, handling three cases:

1. Same itemsize — reshape through uint8 and convert back
2. Larger → smaller dtype — expand last dimension
3. Smaller → larger dtype — collapse last dimension

I'm using the TF backend (lines 999–1043) and PR keras-team/keras#20883 as references. The changes will go into keras/src/backend/openvino/numpy.py and update excluded_concrete_tests.txt.

I'll have a PR up within the next few days. Should I tag @rkazants for review, or would you prefer to review it yourself?

[rishi-sangare]

Update: I've submitted PR keras-team/keras#22407 implementing numpy.view for the OpenVINO backend. All CI checks are passing.

The implementation uses three execution paths:

1. Constant folding — parameter-free subgraphs (constants, broadcasts, etc.) are evaluated on CPU and the raw bytes are reinterpreted via ov.Tensor
2. Symbolic bitwise decomposition — for integer↔integer views on non-constant tensors, shift/mask/OR operations preserve the bit pattern in the OV graph
3. NumPy fast path — plain np.ndarray inputs are handled directly

This also supports bfloat16 reinterpretation through raw byte manipulation (avoiding numpy dtype limitations).

The corresponding test_view exclusions in excluded_concrete_tests.txt have been removed.

@rkazants would appreciate your review when you get a chance!