LLM training in simple, raw C/CUDA

Flash Attention in ~100 lines of CUDA (forward pass only)

Reference implementation of Megalodon 7B model

Lightweight Llama 3 8B Inference Engine in CUDA C