English

This release fixes the most failure-prone part of starting up: the kernel is no longer judged by a fixed stopwatch. Previously the app gave the management port 10 seconds after launch, so machines with many plugins were reported as a failed start before they finished loading, and updating the kernel could even roll a healthy new version back as "failed to start". The app now keeps waiting for as long as the kernel process is alive and reports readiness on its own; an exited process, application shutdown, or a listening management port are the only three outcomes. Request details also gain a choice of how output speed is calculated, between end-to-end throughput and pure generation time with first-token latency removed.

If you hit problems, tell us on Discord or open an issue on GitHub.

Fixed

• Kernel startup is no longer cut off by a fixed timeout: the readiness wait was hardcoded to 10 seconds, so machines with several plugins and a slow first load saw a failed start. The wait now ends only when the kernel process exits, the application is shutting down, or the management port accepts connections; a live process is simply waited for, and slow starts are no longer misreported as failures.
• Updating the kernel no longer rolls back on a slow start: upgrades and the Home start button share one startup path, where a slow boot previously surfaced as "the new version failed to start" and restored the old one. A slow start now just takes longer instead of reverting an install that was already in place.

Added

• Request details can switch how speed is calculated: tokens per second can now be measured over total elapsed time (end-to-end throughput) or with first-token latency removed (pure generation), selectable in the display settings with the list recalculating immediately.

简体中文

这个版本处理启动时最容易出问题的一环：内核启动不再按固定时长判断成败。以前启动后只给 10 秒等管理端口，插件装得多的机器还没加载完就被判为启动失败，更新内核时还会被当成「新版本起不来」自动回滚。现在内核只要还在运行就一直等，直到它自己报告就绪；进程退出、应用关闭或端口就绪是唯一的三种结果。请求明细同时支持切换输出速度的计算方式，可在端到端吞吐率与扣除首字延迟的纯生成阶段之间选择。

如果你在使用中遇到问题，请通过 QQ群 或 GitHub Issues 反馈，我们会尽快处理。

修复

• 内核启动不再被固定超时打断：启动等待原先硬编码为 10 秒，插件较多、首次加载较慢的机器会直接看到启动失败。现在只在内核进程退出、应用正在关闭或管理端口就绪时结束等待，进程还活着就继续等，慢启动不会再被误判。
• 更新内核不会因启动慢而回滚：升级和首页启动走同一条启动路径，之前慢启动会触发「新版本启动失败并回滚旧版本」。现在启动慢只是需要多等一会，不会再把已经装好的新版本退回去。

新功能

• 请求明细可切换速度计算方式：每秒 Token 数现在可以按总耗时（端到端吞吐率）或扣除首字延迟（纯生成阶段）计算，在显示设置里选择，切换后列表即时重新计算。