A new Speculative decoding drafter for vLLM and SGLang called EAGLE-4

EAGLE-4, a speculative decoding drafter released with the eagle-harness package, serves Qwen3.8-27B 3.40× faster than SGLang without speculation for a single request on four NVIDIA B300 GPUs, according to the EAGLE-4 paper and support matrix. The same drafter adapts its draft tree to batch size, hardware and workload while preserving the target model's output distribution, and also runs in vLLM (2.99× at one request) and experimental TokenSpeed (2.02× at 8 concurrent requests); on one AMD MI355X, SGLang with EAGLE-4 is 2.07× faster at 8 concurrent requests. Speedups fall as concurrency rises, to 1.35× in SGLang at 64 concurrent requests and 1.16× in vLLM.

One Flexible Drafter for All Batch Sizes and Hardware Paper EAGLE-4 https://github.com/eagle-hq/EAGLE-4/blob/main/EAGLE4 Paper.pdf · EAGLE-4 draft for Qwen3.8-27B https://huggingface.co/EAGLE-Corp/Qwen3.8-27B-EAGLE4 · Documentation <sub Figure 1 a of the EAGLE-4 paper: Qwen3.8-27B on the SPEED-Bench qualitative split, 4× NVIDIA B300, one request, temperature 0. The numbers this release measured with EAGLE-4 are in the support matrix https://github.com/eagle-hq/EAGLE-4/blob/main/docs/support-matrix.md .</sub EAGLE-4 Extrapolation Algorithm for Greater Language-model Efficiency speeds up LLM inference with speculative decoding. One draft model adapts its draft tree to the batch size, the hardware and the workload, and keeps the target model's output distribution. On four B300 GPUs it serves Qwen3.8-27B 3.40× faster than SGLang without speculation for one request at a time. This repository is eagle-harness , the package that runs EAGLE-4 drafts inside SGLang, vLLM and TokenSpeed. You need Linux, four NVIDIA GPUs with a CUDA 13 driver tested on B300 and about 70 GB of disk for Qwen3.8-27B https://huggingface.co/Qwen/Qwen3.8-27B and its EAGLE-4 draft https://huggingface.co/EAGLE-Corp/Qwen3.8-27B-EAGLE4 , which download on the first start. Start an SGLang server with Docker: docker run --gpus all --ipc host -p 127.0.0.1:30000:30000 \ -e HF HOME=/hf -v ~/.cache/huggingface:/hf \ ghcr.io/eagle-hq/eagle-4-sglang:0.1.0 \ sglang serve --model-path Qwen/Qwen3.8-27B --tp-size 4 --host 0.0.0.0 \ --disable-overlap-schedule --speculative-algorithm EAGLE HARNESS \ --speculative-draft-model-path EAGLE-Corp/Qwen3.8-27B-EAGLE4 \ --speculative-num-steps 12 --speculative-eagle-topk 4 --speculative-num-draft-tokens 16 \ --speculative-algorithm-config '{}' Or install with pip, in Python 3.12. Pip installs also need the CUDA toolkit and a C++ compiler https://github.com/eagle-hq/EAGLE-4/blob/main/docs/install.md requirements : python3.12 -m venv .venv && source .venv/bin/activate pip install "eagle-harness sglang " "sglang==0.5.23.dev252+eagle.5" \ --extra-index-url https://eagle-hq.github.io/EAGLE-4/whl/cu130/ \ --extra-index-url https://download.pytorch.org/whl/cu130 \ --extra-index-url https://pypi.nvidia.com sglang serve --model-path Qwen/Qwen3.8-27B --tp-size 4 \ --speculative-algorithm EAGLE HARNESS \ --speculative-draft-model-path EAGLE-Corp/Qwen3.8-27B-EAGLE4 When the server logs The server is fired up and ready to roll , send a request from another terminal: curl http://localhost:30000/v1/chat/completions -H 'Content-Type: application/json' \ -d '{"model": "Qwen/Qwen3.8-27B", "messages": {"role": "user", "content": "Explain how a hash table works."} , "max tokens": 256}' For vLLM, TokenSpeed experimental and AMD MI355X, follow the engine guides more . Each speedup divides EAGLE-4 throughput by the same engine's throughput without speculation, on Qwen3.8-27B at temperature 0 on the SPEED-Bench client, with the measured recipe of each engine guide. The support matrix https://github.com/eagle-hq/EAGLE-4/blob/main/docs/support-matrix.md has the tokens/s, the correctness checks and how each number was measured. | Engine | GPUs | 1 request | 8 concurrent | 64 concurrent | |---|---|---|---|---| | SGLang | 4× B300 | 3.40× | 1.91× | 1.35× | | vLLM | 4× B300 | 2.99× | 1.70× | 1.16× | | TokenSpeed, experimental | 4× B300 | not measured | 2.02× | 1.42× | On one AMD MI355X, SGLang with EAGLE-4 is 2.07× faster at 8 concurrent requests and 1.56× at 32. - Install https://github.com/eagle-hq/EAGLE-4/blob/main/docs/install.md : pins, wheel indexes, Docker images and the ROCm closure. - Engine guides with the measured recipes: SGLang https://github.com/eagle-hq/EAGLE-4/blob/main/docs/quickstart-sglang.md , vLLM https://github.com/eagle-hq/EAGLE-4/blob/main/docs/quickstart-vllm.md , TokenSpeed, experimental https://github.com/eagle-hq/EAGLE-4/blob/main/docs/quickstart-tokenspeed.md , AMD MI355X https://github.com/eagle-hq/EAGLE-4/blob/main/docs/quickstart-amd.md . - Support matrix https://github.com/eagle-hq/EAGLE-4/blob/main/docs/support-matrix.md and known limits https://github.com/eagle-hq/EAGLE-4/blob/main/docs/known-limits.md . - Calibration https://github.com/eagle-hq/EAGLE-4/blob/main/docs/calibration.md , configuration https://github.com/eagle-hq/EAGLE-4/blob/main/docs/configuration.md and troubleshooting https://github.com/eagle-hq/EAGLE-4/blob/main/docs/troubleshooting.md . - How it works https://github.com/eagle-hq/EAGLE-4/blob/main/docs/architecture.md , kernels https://github.com/eagle-hq/EAGLE-4/blob/main/docs/kernels.md , compatibility https://github.com/eagle-hq/EAGLE-4/blob/main/docs/compatibility.md , new models https://github.com/eagle-hq/EAGLE-4/blob/main/docs/new-model.md and the FAQ https://github.com/eagle-hq/EAGLE-4/blob/main/docs/faq.md . If you use EAGLE-4, cite the EAGLE-4 paper. EAGLE-4 builds on EAGLE, EAGLE-2 and EAGLE-3, which CITATION.cff https://github.com/eagle-hq/EAGLE-4/blob/main/CITATION.cff also lists. @misc{eagle4, author = {EAGLE-4 Team}, title = {{EAGLE-4}: One Flexible Drafter for All Batch Sizes and Hardware}, year = {2026}, note = {Preprint} } - Thanks to Aditya Nandakumar @proaditya and the contributors to the original EAGLE-4 PR 1 for the ROCm port of EAGLE-4 on AMD Instinct MI355X. - Thanks to Andy Luo @andyluo7 and the contributors to the original EAGLE-4 PR 2 for the AMD AITER and Triton kernels on GFX950. - Thanks to Andy Luo @andyluo7 , Aditya Nandakumar @proaditya and the contributors to the merged AMD series in the original EAGLE-4 PR 4 for AITER/FA3 and Split-KV on GFX950. - Thanks to Nicolas Perrin @nperrin-fr and the TokenSpeed/LightSeek contributors for the TokenSpeed implementation in the original EAGLE-4 PR 3, including the MLP L2 warm-ahead and TGV SwiGLU work used in the TokenSpeed fork. - SPEED-Bench is by NVIDIA. Model weights are licensed by their owners. Eagle 4 is source-available software, not OSI-approved open-source software . The terms of the Eagle 4 Source-Available License https://github.com/eagle-hq/EAGLE-4/blob/main/LICENSE govern its use, modification, and distribution, and incorporate the Eagle Trademark Policy https://github.com/eagle-hq/EAGLE-4/blob/main/TRADEMARKS.md . See also the NOTICE https://github.com/eagle-hq/EAGLE-4/blob/main/NOTICE file. Commercial Use of a Competing Offering requires a separate written commercial license regardless of revenue. Other Commercial Use is permitted only within the conditions of the license, including the US $20 million Aggregate Revenue threshold measured over any consecutive 12-month period. If the threshold is exceeded, a separate commercial agreement is required before beginning or continuing any Commercial Use. The license specifies that the requirement continues after a subsequent revenue decrease. Downloading, cloning, forking, copying, accessing, modifying, using, hosting, distributing, or otherwise exercising any licensed right constitutes acceptance of the Source-Available License and incorporated Eagle Trademark Policy. Mere viewing or inspection does not itself constitute trademark use. For separate commercial terms, see the Commercial License Agreement https://github.com/eagle-hq/EAGLE-4/blob/main/COMMERCIAL-LICENSE-AGREEMENT.md and contact licensing@eagle.ai mailto:licensing@eagle.ai . Publishing that agreement does not grant commercial-license rights; it becomes binding only as stated in its Section 1. Files derived from sglang, vLLM, FlashInfer and other projects keep those projects' notices, listed in THIRD PARTY NOTICES.md https://github.com/eagle-hq/EAGLE-4/blob/main/THIRD PARTY NOTICES.md . By submitting, uploading, pushing, committing, proposing, or otherwise providing a pull request, patch, issue attachment, model, weight, dataset, documentation, or other contribution to this repository, you agree to the Eagle Contribution License and Submission Terms https://github.com/eagle-hq/EAGLE-4/blob/main/docs/legal/Eagle-Contribution-Terms-v1.0.pdf PDF . If you do not agree, do not submit a contribution. If you act for an employer, client, partner or other entity, you represent that you have authority to submit the contribution and have obtained all required permissions. Submitting a contribution does not change the license of existing Eagle code; Eagle AI Corp controls outbound licensing for accepted contributions. See CONTRIBUTING.md https://github.com/eagle-hq/EAGLE-4/blob/main/CONTRIBUTING.md for the licensing structure and for how to build, test and open a pull request.