SmoothQuant is constructed before IgnoredScope is applied, so no config can avoid the AttributeError on weight-sharing transformers

Summary

nncf.quantize(..., model_type=ModelType.TRANSFORMER) raises AttributeError inside SmoothQuant's get_weight_value on an RT-DETRv2 ONNX export, and no IgnoredScope a caller supplies can avoid it, because SmoothQuant is constructed before the ignored scope is applied.

Affects: 3.2.0 through develop (3.3.0 does not fix it).

Root cause, from the source

In src/nncf/quantization/algorithms/post_training/pipeline.py, SmoothQuant is appended as the first pipeline step (lines 81-99) whenever model_type == TRANSFORMER and smooth-quant alpha is >= 0 — and it is constructed without ignored_scope. That argument is passed only to the subsequent MinMax step (lines 113-114).

In the OpenVINO backend, get_weight_value → get_weight_tensor_port_id → node.layer_attributes.get_const_port_ids() returns None for a weight shared across decoder layers via Identity nodes — which is exactly how RT-DETRv2 shares its query_pos_head MLP.

Why it is not merely cosmetic

model_type=TRANSFORMER is the single most effective lever we found for transformer detectors: it recovered RT-DETRv2 INT8 from 0.4921 to 0.6680 mAP on host, and to 0.6744 on Tiger Lake. The crash forces every user of a weight-sharing transformer to disable SmoothQuant wholesale to get any of that benefit.

Workaround

AdvancedQuantizationParameters(
    smooth_quant_alphas=AdvancedSmoothQuantParameters(convolution=-1, matmul=-1)
)

This makes the >= 0 guard false and drops SmoothQuant from the pipeline entirely.

Suggested fix

Pass ignored_scope into the SmoothQuant constructor, or defer its construction until after the ignored-scope resolution that MinMax already receives.

Verified

SmoothQuant's openvino_backend.py and algorithm.py are byte-identical (md5) across v3.2.0, v3.3.0 and develop; pipeline.py is byte-identical between 3.2.0 and 3.3.0.