Context

Matmul operation in OpenVINO assumes an implicit shape alignment for input arguments. It applies transpositions specified by optional transpose_a and transpose_b attributes: OV spec.
Currently, weight compression in NNCF does not support transpose_a=True.
Here's the check and test.
Potentially, it affects Mixed-Precision, AWQ, Scale Estimation and Lora Correction algorithms.

What needs to be done?

The task is to enable data-aware weight compression methods (Mixed-Precision, AWQ, Scale Estimation, Lora Correction) for models with transposed input matrix multiplications.

1. At least one function process_stats should be corrected, check - removed.
2. test should pass and be a templated one in order to check OV and Torch backend at once.
3. Tests that used LMLinearModel with transpose_a=False by default should pass with transpose_a=True.

Example Pull Requests

#3179
#3129

Resources

• Contribution guide - start here!
• Intel DevHub Discord channel - engage in discussions, ask questions and talk to OpenVINO developers
• How to link your Pull Request to an issue

Contact points

@ljaljushkin

Ticket

No response