Official Inference Code of EAGLE-4
type: explanation owner: eagle-harness last_verified: 2026-10-21
EAGLE-4 documentation
To serve a model with EAGLE-4, install one engine and follow its guide. SGLang needs no preparation, vLLM reaches its measured speed after a one-time calibration, and TokenSpeed runs on NVIDIA GPUs.
| Task | Guide |
|---|---|
| Install an engine with pip, or pick a Docker image | Install |
| Serve on SGLang | SGLang guide |
| Serve on vLLM | vLLM guide |
| Serve on TokenSpeed | TokenSpeed guide |
| Serve on AMD MI355X | AMD guide |
| See measured speed and correctness | Support matrix |
| Check unsupported or unmeasured behavior | Known limits |
| Measure and fit serving costs | Calibration |
| Look up config keys and defaults | Configuration |
| Diagnose install, startup and generation errors | Troubleshooting |
| Train, export and qualify a draft for another target | New-model onboarding |
| Understand plugin hooks and decode rounds | Architecture |
| Check kernel implementations and fallbacks | Kernels |
| Check host versions and release numbering | Compatibility |
| Convert an existing EAGLE-4 configuration | Migration |
| Find short answers | FAQ |