mdl-latency-test

Readme

# dsh-model-latency-test

**模型测试：一键测出每个模型的首 token 延迟（TTFT）。**
**Model Test: measure first-token latency (TTFT) for every model you have configured.**

> 仓库名 `dsh-model-latency-test` = 包名 `mdl-latency-test`（安装后以包名为准）。
> Repository `dsh-model-latency-test` ships the package `mdl-latency-test`.

---

## 中文说明

这是个 DeepSeek Harness 插件，在**设置**里加一页「**模型测试**」，用来比较你所有已配置模型提供方里每个模型的实际响应速度。

**怎么用**

1. 打开 设置 → 模型测试
2. 每个提供方是一张可折叠卡片，卡片头部有「测试」按钮
3. 点「测试」，它按顺序把该提供方下的每个模型各跑一次，实时显示 `xxx ms` 或「失败」

**它测什么**

- 提供方列表：`llm.listProviders()`
- 模型列表：`llm.listModels(provider)`
- 延迟口径：`llm.stream(...)` 里**第一个非空 token** 的时间差，也就是 TTFT（首 token 延迟），单位毫秒
- 失败的模型单独标「失败」，不影响后面的模型继续测

**注意**

- 测试是串行的（一个一个来），一个提供方模型多的时候需要等一会儿
- 每次测试都会真的发一次最小请求，会消耗一点 token
- 测出来的是"你这条网络 + 这个提供方"当时的真实延迟，跨天跨时段对比更有意义

---

## English

A DeepSeek Harness plugin that adds a **Model Test** page to **Settings**, so you can compare how fast the models you actually have configured respond.

**Usage**

1. Open Settings → 模型测试 (Model Test)
2. Each provider is a collapsible card with a **测试 (Test)** button in its header
3. Pressing it runs that provider's models one after another and shows `xxx ms` or a failure per model, live

**What it measures**

- Providers: `llm.listProviders()`
- Models: `llm.listModels(provider)`
- Latency: the time to the **first non-empty token** of `llm.stream(...)` — TTFT, in milliseconds
- A failing model is marked as failed and does not stop the rest

**Notes**

- Tests run serially, so a provider with many models takes a while
- Each test sends one minimal real request, so it spends a few tokens
- The number is the real latency of your network plus that provider at that moment; compare over time rather than once

---

## Install / 安装

* Through the DSH plugin market, once listed. / 收录后可在插件市场一键安装。
* From GitHub / 从 GitHub 安装:
  `dsh plugin add github:Mo3ter/dsh-model-latency-test`
* Locally from a checkout / 本地目录安装:
  `dsh plugin add file:./dsh-model-latency-test`

Host half: two same-origin HTTP routes under `/plugins/mdl-latency-test/api`
(`GET /list`, `POST /test`) reached through `ctx.llm` and `ctx.webServer`.
Client half: a `settings.section` entry; it needs the Settings UI, so it shows up
in the DSH Web GUI and in DSH Desktop.

需要设置页（Settings UI），所以 DSH Web GUI 与 DSH Desktop 里都能用；没有客户端 UI 的纯命令行环境看不到这一页。

## License

MIT