Yeah, I just reproduced the GPT2 from scratch results in 8.75 hours on 4x 7900 XTX. The fork is here: https://github.com/anthonix/llm.c