Implementation of GPT-2 in pure CMake
Summary
The repository implements GPT-2 entirely in pure CMake using 16.16 fixed-point arithmetic, offering both a toy model and a pathway to a full model via safetensors. It provides scripts to fetch the model artifacts (model.safetensors, vocab.json, merges.txt) from HuggingFace and to generate builds with gpt2_full.cmake or a basic gpt2.cmake workflow. The project showcases an unconventional approach to AI model deployment by pushing more logic into CMake and raises questions about practicality, performance, and portability of such a setup.