Achieve ~2x speed-up in LLM inference with Medusa-1 on Amazon SageMaker AI Researchers developed Medusa, a framework to speed up LLM inference by adding extra heads to predict multiple tokens simultaneously. This post demonstrates how to use Medusa-1, the first version of the framework, to speed up an LLM by fine-tuning it on Amazon SageMaker AI and confirms the speed up with deployment and a simple load test. Medusa-1 achieves an inference speedup of around two times without sacrificing model quality, with the exact improvement varying based on model size and data used. In this post, we demonstrate its effectiveness with a 1.8 times speedup observed on a sample dataset.
https://aws.amazon.com/blogs/m....achine-learning/achi