[NNCF] Enable data-aware weight compression for MatMul with transpose_b=False

[Varshith-Yadav]

Changes

fixes #3494
Added full support for data-aware weight compression when MatMul nodes use transpose_b=False.
Updated and validated test_compression_with_transpose to ensure it passes for transpose_b=False.

Reason for changes

Previously, NNCF’s weight compression flow assumed that the weight input of MatMul operations was always transposed (transpose_b=True).

Related tickets

Tests

pytest tests/openvino/native/quantization/test_weights_compression.py -v
(All tests pass; test_scale_estimation[True] remains the expected XFAIL for ticket 176465.)

[ljaljushkin]

@Varshith-Yadav, thank you for the contribution!
Is it possible to avoid many if conditions and do transpose once to keep original logic for transposed weights?

[ljaljushkin]

please also add unit tests.
at least you can copy-paste from #3725: https://github.com/openvinotoolkit/nncf/pull/3725/files#diff-223ea638f7751f7c0c3e8f867ec9c8c132a3ccd62a9dcea2a5d158836c71c222R1955-R1979 and make sure exception is not raised for transpose_b=False

[ljaljushkin]

@Varshith-Yadav, thank you for the contribution! Is it possible to avoid many if conditions and do transpose once to keep original logic for transposed weights?

I have reconsidered and now believe that transposing each weight can extend the total compression duration. What about implementing and utilizing a "slice_weight" method with a transpose parameter?

[Varshith-Yadav]

@ljaljushkin
That makes sense. I agree that explicitly transposing the full weight tensor could introduce unnecessary overhead

I will update the implementation to use a slice_weight helper method . This way, we can fetch the necessary channels dynamically based on the transpose_b parameter without physically reshaping the underlying tensor.

I'll proceed with this approach and update the PR shortly.

[Varshith-Yadav]

@ljaljushkin
I've updated the implementation as requested. I added a slice_weight helper in utils.py to handle the data access without performing a full transpose, and refactored the GPTQ logic to use it.

I also added a new test file test_utils_slice_weight.py to verify the helper works correctly for both Numpy and PyTorch tensors with different transpose_b settings.

[ljaljushkin]

I believe you need just 2 methods

def get_weight_slice(weight: Tensor, slice_obj: Union[int, slice, Tensor], is_transposed: bool) -> Tensor:
     return weight[:, slice_obj] if is_transposed else weight[slice_obj, :]

def set_weight_slice(weight: Tensor, slice_obj: Union[int, slice, Tensor], value: Tensor, is_transposed: bool) -> None:
    if is_transposed:
        weight[:, slice_obj] = value
    else:
        weight[slice_obj, :] = value

[ljaljushkin]

the same issue should be in other data-aware algorithms: awq, lora_correction, scale_estimation
Support copy-pasting a test for transpose_b=False + all these methods and check whether it fails: https://github.com/openvinotoolkit/nncf/pull/3725/files#diff-223ea638f7751f7c0c3e8f867ec9c8c132a3ccd62a9dcea2a5d158836c71c222R1960-R1961

[ljaljushkin]

utils name violates the code style: https://github.com/openvinotoolkit/nncf/blob/develop/docs/styleguide/PyGuide.md#474-file-naming
Possible name: tensor_slicing.py

[ljaljushkin]

also recommend configuring automatic code formating: https://github.com/openvinotoolkit/nncf/blob/develop/docs/styleguide/PyGuide.md#2-automating-code-formatting

[Varshith-Yadav]

@ljaljushkin Thanks for the detailed feedback! I have updated the PR with the following changes:

Refactored tensor_slicing: I implemented the simplified get_weight_slice and set_weight_slice helpers exactly as suggested (using generic slicing) in src/nncf/quantization/algorithms/weight_compression/tensor_slicing.py.

Algorithm Support: I updated GPTQ, AWQ, Scale Estimation, and LoRA Correction to use these new helpers. They now correctly identify the reduction axis based on transpose_b to handle non-transposed weights.

Testing:

Added test_compress_weights_algorithms_transpose_b_false which successfully verifies that all 4 algorithms work on a model with transpose_b=False without crashing.

Added a new test file tests/openvino/native/test_weight_compression_utils.py to unit-test the helpers with both Numpy and PyTorch tensors.

Formatting: Ran pre-commit to apply the automatic Ruff formatting.

Ready for review!

[ljaljushkin]

Thanks @Varshith-Yadav!
@daniil-lyakhov, could you please help with the review?

[daniil-lyakhov]

@Varshith-Yadav, thank you for your contribution!
My initial comments are below. Besides that, I believe we have to expand tests for each compression algorithm with the transpose_b option for each backend. I'm working on a similar issue right now (support of transpose_a), and when I'll finish my tests I'll share them with you so you can do the same in your PR.

Thank you!

[daniil-lyakhov]

I'm not sure we really need to make a separate function for that. An example on how to do slicing using the build in slice: https://github.com/openvinotoolkit/nncf/pull/3725/files#diff-cefaf6a4a2cb473c23106efa01889f05dc899e43c0dfc74ef8e8d60830e8a467R276-R281

[daniil-lyakhov]

Looks like an openvino/onnx specific code, please introduce a backend method to get this value from the model

[daniil-lyakhov]

Openvino/ONNX specific code as well

[daniil-lyakhov]

Openvino/ONNX specific code in a common algorithm

[daniil-lyakhov]

Onnx/Openvino specific code in a common algorithm

[Varshith-Yadav]

@daniil-lyakhov Thanks for the review!

I have removed the specific OpenVINO attribute checks from the common algorithms. I introduced get_weight_transpose_b in the Backend interface and implemented it in all backends (openvino_backend.py, onnx_backend.py, torch_backend.py, torch_fx_backend.py).

I removed the custom utils.py and switched to using standard Python slice() objects and inline if/else checks as suggested.

I've kept my current integration test (test_compress_weights_algorithms_transpose_b_false in test_utils_slice_weight.py) included for now to verify the fix works. I am happy to update/replace this with your standardized test pattern once you share it.