Megalodon: Efficient LLM Pretraining and Inference with Unlimited Context Length

https://arxiv.org/abs/2404.08801