Design goal
low-latencyresearchproduction
Notable features
Feature-level speculative decoding: instead of a separate draft LM, EAGLE trains a lightweight autoregressive head over the target model's own top-layer features. EAGLE-1 reframed speculative sampling as a feature-uncertainty problem (arXiv:2401.15077); EAGLE-2 uses dynamic draft trees (arXiv:2406.16858); EAGLE-3 removes the feature-prediction constraint and uses training-time test (arXiv:2503.01840, NeurIPS'25)EAGLE-3 is lossless in the paper's terms - it does not change the target model's output distributionComposable with other acceleration work: the README lists vLLM, DeepSpeed, Mamba, FlashAttention, quantization and hardware optimisation as combinableThe README reports a 2x speedup on gpt-fast as the headlineSupport spans vLLM (PR 16937), SGLang speculative decoding, vllm-project/speculators and Intel LLM Library for PyTorch - i.e. the paper's implementation is now consumed by other projectsThe README recommends SGLang's SpecForge for out-of-the-box EAGLE-3 training as of 2025-07-23, which is the clearest signal that the mainstream path moved elsewhere2546 stars, last push 2026-02-20, 283 open issues
Hardware caveats
LICENSE IS NOT A STANDARD OSI LICENSE. The GitHub API reports spdx_id NOASSERTION and license name 'Other' for this repo. Do not assume permissive terms for commercial use - read LICENSE directly. license is left null here rather than guessedRequires a trained EAGLE head per model and per draft length, so it is not a drop-in for stock checkpoints. EAGLE-3 heads for specific models are published separately and version-lock to the base modelCUDA in practice: a pure-Python repo (874126 bytes of Python and nothing else per the language breakdown) that depends on the host framework and its kernels. No ROCm, Metal or Vulkan path documentedNo GitHub releases at all, so there is no versioned artefact to pin - you install from a commitThe project itself now points elsewhere for the mainstream workflow (SpecForge with SGLang, vllm-project/speculators with vLLM). This repo is the paper implementation and reference training code, not the recommended deployment pathAcceptance rate is the whole game: a 2x claim is a vendor claim on the author's own models and hardware with no methodology table in the README. Speedups shrink on small models and on high-temperature samplingNOT deprecated, unlike [[niche-medusa]] - last push 2026-02-20 and it is the live reference for a technique the field adopted. But it is a reference implementation now, not a product
Best for
Reading the reference implementation of the speculative-decoding technique that the field actually standardised onUnderstanding why feature-level drafting beats token-level drafting, and what training-time test changes about the loss surfaceTraining EAGLE-3 heads with SpecForge + SGLang, or serving them through vLLM's speculators plugin
Avoid for
Production deployment without reading the LICENSE - NOASSERTION is a real risk, not a formalityDrop-in speedups on stock checkpoints, which need per-model trained headsNon-CUDA hosts, and any workflow that needs a supported release rather than a commit
notes
The speculative-decoding record that survived. Contrast with [[niche-medusa]]: EAGLE kept being implemented on (last push 2026-02-20 vs Medusa's 2024-06-25), produced the three-paper lineage that the serving stacks cite, and got absorbed into vLLM, SGLang, vllm-project/speculators and Intel's ip-llm. arXiv titles/dates for 2401.15077, 2406.16858, 2503.01840 and the NeurIPS'25 acceptance all verified via the arXiv API; all repo metadata from the GitHub API on 2026-10-03. ||| MAINTENANCE (dated fact, GitHub API, 2026-10-04): last commit 2026-02-20 (7.4 months before this record), NO GitHub release or tag found, 28 contributors (distinct committers including anonymous), open_issues_count 101 (issues PLUS pull requests - see [[ev-github-open-issues-count-includes-pull-requests]]), not archived, 2546 stars. See [[ev-maintenance-status-is-four-dated-facts]] for how to read this fact set. ||| DEPLOYMENT: NO EVIDENCE OF PRODUCTION DEPLOYMENT FOUND as of 2026-10-04 - repositories, READMEs, docs and issue trackers were searched. See [[ev-production-evidence-is-scarce-and-nearly-always-single-site]] for the grading scheme; note that for an artifact-evaluation or paper-code release, non-deployment is documented intent rather than a negative finding.