Introduction

In the development and debugging of model inference, merely obtaining average latency is often insufficient to reveal system performance bottlenecks. In scenarios requiring precise evaluation of response quality, calculating latency percentiles is a necessary approach. DeepSeek Harness (DSH) adopts a plugin-based architecture and supports extending capabilities through community plugins. The plugin introduced below is dsh-latency-pct, a plugin focused on latency percentile calculation.

What It Is

dsh-latency-pct is a DSH plugin maintained by uckkk. Its core value lies in accurately calculating latency percentiles by parameter and returning detailed results. The plugin is implemented in pure Node.js, does not depend on a network environment, and can be run locally.

Core Features

  1. Latency percentile calculation: supports precise calculation by parameter.
  2. Detailed result output: outputs detailed data after calculation is completed.
  3. No network dependency: implemented in pure Node.js and can run without internet access.

Installation and Activation

Run the following command in the terminal to install the plugin:

dsh plugin add github:uckkk/dsh-latency-pct

After installation is complete, the plugin registers the corresponding tools in the current DSH session.

Typical Usage

After installing and activating the plugin in the session, you can use it by directly calling the tool registered by the plugin. For the exact invocation method, refer to the list of available tools in the current session.

Applicable Scenarios and Notes

  • Applicable scenarios: development and debugging scenarios that require quantitative analysis of model inference latency distribution.
  • Dependency requirements: before installation, ensure the environment satisfies the following peerDependencies:
  • @deepseek-ai/cordis >=4.0.1
  • @deepseek-ai/dsh-tools >=0.1.0-rc.6
  • Permissions note: the plugin runs with the permissions of the current DSH process. It is recommended to review the source code and license (MIT) before use.

Conclusion

By following the steps above, you can obtain precise latency percentile data in a local environment. For more details, refer to the plugin directory or source repository.
- Plugin directory: https://www.skillhub.cn/plugins/uckkk/dsh-latency-pct
- Source repository: https://github.com/uckkk/dsh-latency-pct