Implementation of Bottleneck Transformer in Pytorch

A from-scratch implementation of a tiny Transformer inference engine in PyTorch, built to study prefill vs decode performance, attention mechanics, and GPU bottlenecks. The project focuses on correctness and profiling-driven optimization rather than training.