Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads

Medusa technical report. Primary source for the multi-head drafting mechanism that the original vLLM/SGLang n-gram and draft-model paths later replaced.