`_get_smooth_quant_param_grid` is broken

[ruro]

🐛 Describe the bug

nncf/nncf/quantization/algorithms/hyperparameter_tuner/param_grid.py

Line 90 in b34d24e

AdvancedSmoothQuantParameters(matmul=alpha_v) for alpha_v in itertools.product(alpha_values)

I think that this line is currently subtly broken.
Right now, _get_smooth_quant_param_grid() returns

{
    'advanced_parameters:smooth_quant_alphas': [
        AdvancedSmoothQuantParameters(convolution=-1, matmul=(0.15,)),
        AdvancedSmoothQuantParameters(convolution=-1, matmul=(0.25,)),
        AdvancedSmoothQuantParameters(convolution=-1, matmul=(0.5,)),
        AdvancedSmoothQuantParameters(convolution=-1, matmul=(0.75,)),
        AdvancedSmoothQuantParameters(convolution=-1, matmul=(0.95,))
    ]
}

Notice that the matmul values are wrapped in an extra tuple instead of just being plain floats like they are supposed to. This is happening because itertools.product(*iterables) will always return tuples of length len(iterables) even when there is only 1 argument.

These tuple values will then cause a crash in create_ptq_pipeline:

File "/.../site-packages/nncf/quantization/algorithms/post_training/pipeline.py", line 97, in create_ptq_pipeline
    if model_type == ModelType.TRANSFORMER and (sq_params.convolution >= 0 or sq_params.matmul >= 0):
TypeError: '>=' not supported between instances of 'tuple' and 'int'

Environment

NNCF 2.17.0
OpenVINO 2025.2.0
Torch 2.7.1
ONNX 1.18.0

OS Ubuntu 22.04.3 LTS
Python 3.10.12
Install PyPI

Minimal Reproducible Example

I don't have a proper MRE right now, but I think that any quantization pipeline with enabled hyperparameter tuning should demonstrate the crash. So, something along the lines of:

nncf.quantize_with_accuracy_control(
    model=openvino_model,
    ...,
    advanced_accuracy_restorer_parameters=nncf.AdvancedAccuracyRestorerParameters(
        tune_hyperparams=True,
        ...,
    ),
)

Also, I think that the issue if fairly obvious from the _get_smooth_quant_param_grid code mentioned above. I was able to fix the issue by adding the missing tuple unpacking

-            AdvancedSmoothQuantParameters(matmul=alpha_v) for alpha_v in itertools.product(alpha_values)
+            AdvancedSmoothQuantParameters(matmul=alpha_v) for alpha_v, in itertools.product(alpha_values)

Are you going to submit a PR?

• Yes I'd like to help by submitting a PR!

[andrey-churkin]

Hi @ruro, thanks for reporting the problem. I will check it and get back to you as soon as possible.