Shift Negative Correction does not work with correctly with activation mixed precision (Max cut computation issue)

[ofirgo]

Issue Type

Bug

Source

source

MCT Version

2.3-dev

OS Platform and Distribution

No response

Python version

No response

Describe the issue

When running MCT with shift negative activation correction, the activation quantization of each relevant layer would automatically change to 16bit (regardless any TPC configuration definitions).

When running activation mixed precision quantization that uses max cut as activation memory metric, all layers that are quantizing activations are included as part of a cut memory, which includes the 16 bit quantized outputs of layers that have been through SNC substitution.
This is not reflected in the resource utilization data computation for activation memory, hence, the results of the mixed precision might be compromised.

Expected behaviour

The SNC 16 bit activations should be considered in the resource utilization data computation or treated differently during mixed precision activation memory estimation.

Code to reproduce the issue

run Yolov8n-pose with SNC and activation MP

Log output

No response