# hao-ai-lab/LookaheadDecoding

[ICML 2024] Break the Sequential Dependency of LLM Inference Using Lookahead Decoding

Repository: https://github.com/hao-ai-lab/LookaheadDecoding
Canonical: https://ross.abutalabs.com/products/lookaheaddecoding
Homepage: https://arxiv.org/abs/2402.02057
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2025-03-06T07:17:19+00:00

## Health v2 (maintenance only)
Score: 31/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 10, release rhythm 35, longevity 72
- inputs: {"age_days": 1016, "days_push": 545, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1342, forks 83 (observed 2026-08-28T04:04:26.323289+00:00)

## What it is
A Python library implementing Lookahead Decoding, an exact parallel decoding algorithm that accelerates LLM inference without a draft model or data store. It uses Jacobi iteration with n-gram caching and verification branches to reduce decoding steps, integrating with FlashAttention and supporting multi-GPU scaling.

## Use cases
- speed up autoregressive LLM text generation
- accelerate code completion with LLMs
- reduce LLM decoding latency without a draft model
- parallelize LLM inference across multiple GPUs
- integrate faster decoding into my own sampling code
- speed up LLaMA-2 chat generation

## When to choose
- you want speculative-decoding-like speedups without training or obtaining a draft model
- you need exact decoding with no quality loss
- you run inference on modern GPUs with FlashAttention support
- you want to reduce decoding steps in code completion workloads

## When to avoid
- you need a production inference server with broad model support out of the box
- your workload is memory-abundant and batch-heavy where parallel decoding gains are minimal
- you require long-term maintenance guarantees or wide ecosystem integration

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, machine-learning, gpu-computing
- domain: large-language-models, machine-learning, deep-learning, gpu-computing
- platform: python
- tags: lookahead-decoding, jacobi-decoding, speculative-decoding, inference-acceleration, flashattention, icml-2024, gpu, linux

## Member repositories
- hao-ai-lab/LookaheadDecoding (main) score 31

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:26.323289+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:43:05.686416+00:00, confidence not recorded.
  - readme: https://github.com/hao-ai-lab/LookaheadDecoding (fetched 2026-08-28T04:04:26.323289+00:00, sha 333916e5d4fe)
  - homepage: https://arxiv.org/abs/2402.02057 (fetched 2026-08-29T12:02:28.396491+00:00, sha 35fa87771f0e)
  - site_page: https://info.arxiv.org/about/donate.html (fetched 2026-08-29T12:02:28.405835+00:00, sha cca9c3a11c56)
  - site_page: https://info.arxiv.org/about/ourmembers.html (fetched 2026-08-29T12:02:28.410152+00:00, sha 47cbc55ff1de)
  - site_page: https://info.arxiv.org/about (fetched 2026-08-29T12:02:28.412231+00:00, sha a1f16f915a9a)
  - site_page: https://info.arxiv.org/labs/index.html (fetched 2026-08-29T12:02:28.407961+00:00, sha b14a8d05a0ec)
- Data as of 2026-08-30T08:39:29.467469+00:00.