MTP Speculative Decoding

Multi-Token Prediction (MTP) Speculative Decoding is a technique for accelerating large language model inference. Its core mechanism involves using a lightweight draft model to predict multiple consecutive tokens in a single step, after which the primary model performs batched parallel verification of those candidate tokens. This reduces the number of autoregressive iterations required by the primary model, significantly improving decoding efficiency while preserving output quality. The technique is commonly applied to reduce inference latency and increase throughput in large models.