Transformer tricks: Precomputing the first layer

arXiv:2402.13388v1 Announce Type: new
Abstract: This short paper describes a trick to speed up inference of transformers with RoPE (such as LLaMA, Mistral, and PaLM). For these models, a large portion of the first transformer layer can be precomputed, which results in slightly lower latency and lower cost-per-token. Because this trick optimizes only one layer, the relative savings depend on the total number of layers. For example, the maximum savings for a model with only 4 layers (such as Whisper …

abstract arxiv cost cs.lg inference latency layer llama mistral palm paper per rope speed token transformer transformers trick tricks type