# SafeAILab/EAGLE

Official Implementation of EAGLE-1 (ICML'24), EAGLE-2 (EMNLP'24), and EAGLE-3 (NeurIPS'25).

Repository: https://github.com/SafeAILab/EAGLE
Canonical: https://ross.abutalabs.com/products/safeailab-eagle
Homepage: https://arxiv.org/pdf/2503.01840
Language: Python
License: NOASSERTION
License Family: other
Topics: large-language-models, llm-inference, speculative-decoding
Last push: 2026-02-20T02:53:29+00:00

## Health v2 (maintenance only)
Score: 57/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 68, release rhythm 35, longevity 71
- inputs: {"age_days": 1000, "days_push": 194, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, no_license
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2510, forks 296 (observed 2026-08-28T04:06:57.494893+00:00)

## What it is
EAGLE is the official implementation of the EAGLE family of speculative decoding algorithms (EAGLE-1/2/3) for accelerating LLM text generation. It uses a lightweight draft model that extrapolates the target LLM's feature vectors to draft tokens, achieving 2-4x lossless speedups over vanilla decoding.

## Use cases
- speed up llm inference
- speculative decoding for large language models
- reduce llm generation latency
- faster text generation without quality loss
- accelerate autoregressive decoding
- deploy llms with lower inference cost

## When to choose
- you need lossless LLM decoding speedups with provable output distribution consistency
- you want the fastest speculative decoding method and can train a draft model on modest GPUs
- you use supported models and want compatibility with vLLM, FlashAttention, or quantization

## When to avoid
- you need plug-and-play inference without training a draft model for your specific LLM
- your model architecture is not among the supported target models
- you only need small-model inference where decoding is not the bottleneck

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, machine-learning, gpu-computing
- domain: large-language-models, machine-learning, deep-learning, performance
- platform: python
- tags: speculative-decoding, llm-acceleration, draft-model, lossless-decoding, inference-optimization, gpu, linux, docker

## Member repositories
- SafeAILab/EAGLE (main) score 57

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:06:57.494893+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 2026-08-30T02:26:23.954381+00:00, confidence not recorded.
  - readme: https://github.com/SafeAILab/EAGLE (fetched 2026-08-28T04:06:57.494893+00:00, sha c2ceb2720eb1)
- Data as of 2026-08-30T08:39:29.467469+00:00.