How to calculate the true used token numbers?

[KaKa-101]

Thanks for your great work.
Could anybody tell me how to calculate the retained token numbers basing the --scale and --bias provided in the eval scripts?

[Gumpest]

Thank you very much for your interest in our work. The computation of the equivalent token count can be found at https://github.com/Gumpest/SparseVLMs/blob/main/llava/model/language_model/modelling_sparse_llama.py#L373.