sgl-project/sglang-jax

[Feature] Multi-Token Prediction (MTP) support for sgl-jax

Open

#192 opened on Sep 17, 2025

View on GitHub
 (3 comments) (1 reaction) (1 assignee)Python (276 stars) (101 forks)auto 404
SpecDecodeenhancementhelp wanted

Description

Motivation

Current autoregressive language models generate tokens sequentially, which creates inherent bottlenecks in inference throughput. While speculative decoding techniques like EAGLE improve performance through draft-verify mechanisms, they still rely on single-token predictions from the base model. Multi-Token Prediction addresses this limitation by enabling the model to directly predict multiple tokens, reducing the number of forward passes required for sequence generation.

Road Map

  • Eagle Worker main process @SiqiLi-Fighting

    • adapted tree mask to rpa v3 attention kernel
    • add bigram key prefix cache to radix cache

  • Performance Optimization @SiqiLi-Fighting

    • eagle topk = 1, JIT functional optimization
    • eagle topk > 1, bulid tree mask kernel at draft decode stage
    • Compatibility with SchedulerOverlap
    • non greedy sampling kernel implement @SiqiLi-Fighting

  • More Speculative algorithms Support (Call for Contribution)

    • ngram algorithms like PLD/Suffix Decoding/ LookAhead adaptation

Contributor guide

Tech stack
python
Domain
machine learningaibackend
Issue type
feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
4
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
over 1 week
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
clear
Prerequisites
PythonJAXMachine Learning
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
30
Research direction
Study the existing Eagle worker implementation in sglang jax, understand the tree mask and attention kernel, then design a multi token prediction mechanism that can predict multiple tokens per forward pass.