Quantized model export_onnx_qop - AssertionError: Output quant required

[shoskensMagics]

I've been playing with the brevitas_examples.imagenet_classification.ptq flow for a little bit. Very impressed on how easy it is to quantize, validate and then export to an interchangeable format like ONNX.
I would like my quantized model to be used further down the line (through TVM on custom hardware, if it matters much).
For this purpose, I expect the ONNX to have QOps, instead of the QCDQ composed nodes.
Instead of calling export_onnx_qcdq, I thought a simple call to export_onnx_qop would do the job.
This does not seem to be the case. I'm getting an AssertionError originating from the place specified below.

│ .../brevitas/export/onnx/standard/qoperator/handler/parameter.py:45 in validate                  │                                                                                                                                          
│                                                                                                  │                                                                                                                                          
│    42 │   @classmethod                                                                           │
│    43 │   def validate(cls, module: QuantWBIOL, requires_quant_bias=True):                       │
│    44 │   │   assert module.is_weight_quant_enabled, 'Weight quant required'                     │
│ ❱  45 │   │   assert module.is_output_quant_enabled, 'Output quant required'                     │
│    46 │   │   # Handling narrow_range is across the network is difficult do to the fact that     │
│    47 │   │   # it's not part of QuantTensor, and so it can't be cached                          │
│    48 │   │   assert not module.is_quant_output_narrow_range, 'Narrow output quant not support   │

Is there something I can do here? I expect my layers to be fully quantizable?
Can I play around with some settings in ptq_common.quantize_model?

[Giuseppe5]

Thanks for you feedback, it is really appreciated!

QOps have more strict requirements compared to QCDQ format.
From the error, it looks like you need to add also an output quantizer to your quantized layer.

We have a notebook tutorial that could be helpful for this issue.

Please let me know if you have further questions.