Laya request serialization, sequence construction and decision decoding follow
the Apache-2.0 Laya SDK by Nandha Kishor / Convai Innovations.
Source: https://github.com/NandhaKishorM/laya
Reference release: laya 0.3.5.
This project implements a separate inference runtime and modifies the request
implementation to use tokenizers and NumPy without Torch or Transformers.
