Show HN: TurboGPT: train 22KiB transformer in 13s
Summary
Show HN: TurboGPT showcases a tiny, byte-level transformer training project implemented in CUDA C++. The GitHub repo documents building, running, and evaluating a 22KiB transformer, with results showing 2.52435 BPB after 1.5B training tokens. The project is public, MIT-licensed, and includes a README and CUDA-based codebase.