Training
EAGLE
Accelerate decodingSpeculative decodingReferenceDraft at the feature (hidden-state) level instead of the token level — higher acceptance than Medusa, the current speculative-decoding SOTA.
- Objective
- Accelerate decoding
- Relieves
- Bandwidth
- Format
- Feature-level autoregressive draft head
- Granularity
- Hidden-state draft + tree verify
- Lifecycle
- QAT / QAD · needs gradients
- Calibration
- Small calibration set
- Compression
- ~2.5–4× decode throughput
- Quality
- 100% — verification keeps only tokens the target model would emit
- Hardware
- Trained EAGLE head; tree-attention verify kernel
- Runtimes
- vLLMSGLang
In the wild
vllm-Moet (kacper-daftcode)
Community vLLM fork with a custom-trained EAGLE-style speculative decoder, built to run a leading open-weights model at home.
AngelSlim Eagle3 training (Tencent)
Turnkey EAGLE-3 draft-model training for Qwen3 / Hunyuan / DeepSeek LLMs, VLMs, and audio models — accept length 1.8–3.5, up to 1.9× decode speedup, exports to vLLM/SGLang.
A worked Spark recipe for this method hasn't been written yet — it lives here as a reference point in the ontology.