Why is qdq_weights=True used when calling sim.export?

[minseop4898]

Hello,

When using aimet_torch, I noticed that in the sim.export function, the code sets the option qdq_weights=True. Could you explain the reason for this? Specifically, in aimet_torch/_base/quantsim.py, inside the export method of the class _QuantizationSimModelBase, the following line appears:

# Create a version of the model without any quantization ops
model_to_export = self.get_original_model(self.model, qdq_weights=False)

This seems to imply that the weights stored in the exported ONNX file might differ from the original model’s weights.
In our case, we would like the exported weights to remain identical to the original model’s weights. Would it be acceptable to simply change this to qdq_weights=True before export?

Thanks in advance!

[quic-mtuttle]

Hi @minseop4898, if I understood correctly you mean to say that inside export, it does

model_to_export = self.get_original_model(self.model, qdq_weights=True)

and you want to change your source code to qdq_weights=False? That should be fine. Doing so would result in the exported model's weights remaining true to the original model's weights without applying any quantize-dequantize operation to them.

The reason for using qdq_weights=True here is to minimize simulation vs on-target mismatches due to slight differences in rounding directions between AIMET quantization and downstream tools like QNN/QAIRT. That said, this difference is usually quite small, so switching to qdq_weights=False is unlikely to cause any issues in most cases.