EAGLE-4 投机解码提速：27B 模型达 863.8 tokens/s
Some of the most useful inference work right now is happening in the speculative decoding component.
EAGLE-4 在 27B 模型单请求下实现 863.8 tokens/s 解码速度，为 DFlash 2 的 1.28 倍、DSpark 的 1.40 倍，测试环境约快 3 倍且输出结果一致。它保留草稿-验证循环，改进集中在草稿侧，采用自适应草稿结构和更廉价的草稿 LM head。Eagle 官方称在 Qwen3.8-27B 上配合 TokenSpeed（BF16、4× NVIDIA B300、TP4、单请求、T=0）可达 961.6 tokens/s。
查看原帖EAGLE-4 is here. From EAGLE-1 to EAGLE-3, we’ve kept making inference faster. EAGLE-4 is our fastest yet. 961.6 tokens/s on @Alibaba_Qwen’s Qwen3.8-27B with TokenSpeed (BF16, 4× @NVIDIA B300, TP4, one request, T=0). Github: https://t.co/GwocZk1UTU https://t.co/GawgFNXKWj
正文
当前显示原文，中文全文译文尚未就绪。
Some of the most useful inference work right now is happening in the speculative decoding component.
It has become one of the most dependable ways to get more tokens per second out of a model you’ve already chosen.
EAGLE-4 decodes 863.8 tokens/s on a 27B model on a single request, 1.28X DFlash 2 and 1.40X DSpark.
EAGLE-4 is a speed-up add-on for running AI chatbots: a tiny helper model guesses the next few words, and the big model checks all those guesses at once instead of writing one word at a time. The answers come out identical, just about 3× faster on the tested setup.
EAGLE-4 keeps the draft-and-verify loop and puts its work into the drafting side, with an adaptive draft structure and a cheaper draft LM head.
The loop:
- A small draft model proposes several tokens ahead - The target model checks all of them in one forward pass - Matching tokens are kept, and the first mismatch is replaced with the target’s own token - Repeat from there