Fix weight layout detection for MatMul with transpose in OpenVINO backend

[naraen-ram]

Fixes #3230

Description

Fix incorrect weight layout detection for MatMul layers when transpose is applied via OpenVINO graph rather than constant attributes.

Previously, constant_layer_attrs["transpose"] did not reflect graph-level transpose nodes, which could lead to incorrect channel axis detection during weight compression.

This change checks input_attributes metadata to correctly determine transpose state before computing layout.

Testing

Reproduced issue using custom Transpose → MatMul OpenVINO model.

Verified using:
pytest tests/openvino/native/quantization/test_weights_compression.py -k matmul

Result:
12 passed, 0 failed

Impact

Fixes weight compression correctness for:

• AWQ
• Mixed precision
• Scale estimation
• LoRA correction

[naraen-ram]

Hi @ljaljushkin ,

Just a follow-up on this PR. I’d appreciate a review whenever convenient.

This contribution is part of my preparation for GSoC 2026, so getting feedback would help me move forward. Happy to make any requested changes.

Thanks!

[daniil-lyakhov]

Hi @naraen-ram, thank you for your contribution.

1. Please describe a use case when this fails a model compression/quantization
2. Layer attributes should not reflect neighboring nodes in any way, the transpose layer attribute is openvino MatMul specific parameter, it does not correlate with anything

[naraen-ram]

Hi @daniil-lyakhov,

During my investigation of the issue, I reproduced a scenario where the transpose appears as a separate Transpose → MatMul pattern in the OpenVINO graph rather than through the transpose_a / transpose_b attributes of the MatMul operator. In this case, the graph contains a Transpose node feeding into MatMul, while the MatMul attributes still report transpose_a=False. My initial implementation attempted to account for this situation during layout detection.

After revisiting the issue description again, I realized that the original task specifically refers to supporting MatMul(transpose_a=True) cases rather than transpose as a separate node.

I'll reproduce a model that produces transpose_a=True in the MatMul attributes and adjust the implementation accordingly.

Thanks!