Skip to content

v0.9.1「Windows 首启修复」/ Windows first-run fix

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 24 Jul 16:17
v0.9.1
036ce08

v0.9.1「Windows 首启修复」/ "Windows first-run fix"

单点修复版:修复 Windows 首次使用时「语音模型下载到 100% 又从头开始、反复下载、就是不出片」的问题(报告自 clipforge#17)。所有 Windows 用户建议升级;macOS 不受此 bug 影响,但也能得到更诚实的下载/解压进度。

🐛 发生了什么

  • 语音模型以 .tar.bz2 归档分发,而 Windows 自带的 tar 只编译了 gzip、不支持 bzip2——模型每次下载完成后,解压都必然失败
  • 旧版把解压失败一律当「下载的文件损坏」处理:删掉刚下完的归档 → 换下一个镜像从零重下 → 进度条再次 0→100% → 再失败……无限循环
  • 于是你看到:进度到 100% 又重头开始、反复多次、网络一直在跑,模型永远装不上,自然不出片

✅ 修了什么

  • 内置解压兜底:系统 tar 失败时自动切换到内置纯 JS 解压器(bz2+tar 全程进程内完成),一次下载即装好——不再依赖任何系统解压能力,Windows/macOS/Linux 行为一致
  • 解压有了自己的进度阶段:「解压安装模型」独立进度条+提示(1GB 模型解压需几分钟,不再看似卡死或循环)
  • 原子落位:解压先进临时目录、成功后一步改名生效——中途退出不再留下会被误判「已安装」的残缺模型
  • 体积提示说真话:下载提示按真实归档大小显示(此前恒为「约 170MB」,SenseVoice 归档实际约 1GB)
  • 升级后,旧版已下载的字节自动续传,流量不浪费

⬆️ 升级方式

到 Releases 下载新安装包覆盖安装即可,已下载的模型与所有本地数据原位保留。

A focused bugfix release: fixes the Windows first-run loop where the speech-model download hit 100%, restarted from zero, and never finished (reported in clipforge#17). All Windows users should upgrade; macOS was never affected by the loop but gains the honest download/extract progress too.

🐛 What happened

  • Speech models ship as .tar.bz2 archives, and Windows' built-in tar is compiled with gzip support only — extraction failed every single time after a complete download
  • Older versions treated any extraction failure as "corrupt download": delete the just-finished archive → re-download from zero via the next mirror → 0→100% again → fail again… forever
  • Hence the symptoms: progress hits 100% and restarts, over and over, network busy the whole time, model never installs, no clips ever produced

✅ What's fixed

  • Built-in extractor fallback: when system tar fails, a pure-JS bz2+tar extractor takes over automatically — one download now installs, with identical behavior on Windows/macOS/Linux
  • Extraction is its own progress stage: unpacking a 1GB model takes a few minutes and no longer looks frozen or looping
  • Atomic install: extraction goes through a staging dir and lands with a single rename — quitting mid-extract can no longer leave a half-installed model that gets mistaken for a working one
  • Honest size hint: the download hint shows the real archive size (previously hard-coded "~170MB"; the SenseVoice archive is actually ~1GB)
  • After upgrading, previously downloaded bytes resume automatically — no bandwidth wasted
Assets 9
Loading