GOODSPEED

Speculative decoding for efficient and fair LLM inference at the edge.

All projects

Efficient inference

GOODSPEED

Speculative decoding for efficient and fair LLM inference at the edge.

  • Speculative decoding
  • Edge inference
  • Fair resource allocation
GOODSPEED distributed speculative decoding architecture with draft and verification servers

Overview

Faster inference, shared fairly.

GOODSPEED explores speculative decoding for efficient large language model inference in distributed edge environments.

The project studies how to balance inference speed, output quality, and proportional fairness across heterogeneous edge resources and draft servers.

Research focus

Coordinated inference at the edge.

  1. Speculative decoding optimization for faster token generation while preserving output quality.
  2. Fair resource allocation across heterogeneous edge clients and servers.
  3. Efficient LLM inference in resource-constrained distributed systems.
  4. Coordinated inference architectures across multiple edge nodes.

System direction

Adaptive decoding across distributed resources.

GOODSPEED builds on recent advances in accelerated inference, distributed machine learning systems, edge computing optimization, and fair resource allocation.

Status

Active research and development.

This project is under active development. Research findings and implementation details will be shared as they become available.

References

2025

  1. GoodSpeed: Optimizing Fair Goodput with Adaptive Speculative Decoding in Distributed Edge Inference
    Phuong Tran, Tzu-Hao Liu, Long Tan Le, and 6 more authors
    2025
    Accepted at INFOCOM 2026