Showing 1 of 1 repositories tagged #flashattention, ranked by stars
GPT-2-style LLM built from scratch in C/CUDA with hand-written backprop, BPE tokenizer, FlashAttention, pretraining, and SFT.
