Symbolic shape inference determines tensor dimensions through a computational graph. In ONNX Runtime static quantization, resolved shapes support calibration; unresolved shapes can produce poor activation ranges and damage accuracy. section 6.4
For BERT, generic ONNX preprocessing failed to complete shape inference, and the resulting static-quantized model reached only 50.92% accuracy. The proposed correction is the transformer-aware optimizer, which resolves shapes before calibration. section 6.4
