Pull requests / #163

#163 feat: show prefill speed alongside decode speed in dashboard

closed · @hendrikp · 0 评论 · 在 GitHub 查看

描述

Adds prefill speed to the existing Speed graph, with independent scaling and a darker curve. Both rates use t/s, with prefill displayed on the right.

Exposes `prefill_tok_s_mean` using the engine’s existing running mean, excluding cached tokens.

<img width="572" height="336" alt="Dashboard showing prefill and decode speed" src="https://github.com/user-attachments/assets/5ba70eaf-3615-4a69-9194-14735c454480" />

Tested with passing server tests and a long prompt + manually.

站内延伸阅读

链到安装、模型与版本说明,便于 SEO/GEO,非官方 issue 正文。