Per-channel zero points but per-tensor scales

[V0XNIHILI]

How would I go about implementing the above feature in Brevitas? Where should I get started and is this even possible in the way Brevitas is currently set up?

[V0XNIHILI]

@Giuseppe5 do you know more about this?

[Giuseppe5]

Yes, apologies for the late reply. Later tomorrow I will publish an example of how to achieve this, but in general it should be possible since zero point and scale have independent shapes that can be individually specified when instantiating a new quantizer.

[V0XNIHILI]

@Giuseppe5 if you have the time, I would still really love to see how we can achieve this with Brevitas 😁

[Giuseppe5]

Apologies for the delay, but it is a bit more complicated than I originally thought.

It is still possible but it requires to play a bit more with dependency injection.
I'll try to look into it but not sure when at this point.
Apologies.

[V0XNIHILI]

No problem! Do you have a reference for me on where to start from/where to look?

[Giuseppe5]

You need to create a quantizer where:

• scaling_shape is scalar shape (ref here:

  brevitas/src/brevitas/quant/solver/parameter.py

  Line 111 in 6bdb1f8

  def scaling_shape(scaling_per_output_channel):

  )
• The quantizer inherits from this class:

  brevitas/src/brevitas/quant/solver/weight.py

  Line 36 in 2004568

  class SolveWeightScalingPerOutputChannelShapeFromModule(ExtendedInjector):

• The zero point shape is equal to: https://github.com/Xilinx/brevitas/blob/200456825f3b4b8db414f2b25b64311f82d3991a/src/brevitas/quant/solver/weight.py#L53C9-L53C41