Karpathy's llm.c trains LLMs in pure C/CUDA, runs 7% faster than PyTorch

The GitHub repository llm.c, led by Andrej Karpathy, is drawing attention as a project that implements large language model pretraining in pure C and CUDA, without relying on heavy frameworks like PyTorch or cPython. Starting from reproducing GPT-2, it now runs roughly 7% faster than PyTorch Nightly on the same training run.