Megalodon: Efficient LLM Pretraining and Inference with Unlimited Context Length
This addresses the bottleneck of quadratic complexity and weak length extrapolation in Transformers for researchers and practitioners needing efficient long-context modeling, representing a strong incremental improvement over existing sub-quadratic methods.
The paper tackles the problem of scaling large language models to long sequences by introducing Megalodon, a neural architecture that achieves a training loss of 1.70 with 7 billion parameters and 2 trillion tokens, outperforming Llama2-7B (1.75) and nearing Llama2-13B (1.67).
The quadratic complexity and weak length extrapolation of Transformers limits their ability to scale to long sequences, and while sub-quadratic solutions like linear attention and state space models exist, they empirically underperform Transformers in pretraining efficiency and downstream task accuracy. We introduce Megalodon, a neural architecture for efficient sequence modeling with unlimited context length. Megalodon inherits the architecture of Mega (exponential moving average with gated attention), and further introduces multiple technical components to improve its capability and stability, including complex exponential moving average (CEMA), timestep normalization layer, normalized attention mechanism and pre-norm with two-hop residual configuration. In a controlled head-to-head comparison with Llama2, Megalodon achieves better efficiency than Transformer in the scale of 7 billion parameters and 2 trillion training tokens. Megalodon reaches a training loss of 1.70, landing mid-way between Llama2-7B (1.75) and 13B (1.67). Code: https://github.com/XuezheMax/megalodon