Max cut algorithm fails to converge on models with MHA

[ofirgo]

Issue Type

Bug

Source

source

MCT Version

Nightly

OS Platform and Distribution

No response

Python version

No response

Describe the issue

When running the max cut scheduling algorithm on a model with a multi-head attention, with num_heads >= 3, the algorithm does not converge to a solution, even if increasing the number of search iteration limits up to 5,000, or increasing the heuristic search boundary.

Expected behaviour

The algorithm should find a max cut solution.

Code to reproduce the issue

Run the max cut algorithm on a model with a single mha layer

Log output

No response

[haihabi]

@ofirgo and @elad-c isnt this issue fixed already?