Skip to content

Speculative Decoding for Faster Large Language Model Inference

Authors: Chen et al. Year: 2023 ArXiv/Link: https://arxiv.org/abs/2302.01318

Summary

Uses a smaller model to predict multiple tokens speculatively, verified by larger model for speedups without quality loss.

Key Concepts

  • Speculative decoding
  • Multi-token prediction
  • Smaller verifier
  • Speedup
  • Token efficiency

Impact

Practical speedup technique for autoregressive generation

Category

Decoding Strategies