[BUG] Windows 桌面端无法启动:GPU 进程初始化失败直接终止应用(Intel Arc + 多虚拟显示器环境) #8255
Replies: 3 comments
|
我也是同样问题,环境不同但现象一致:
错误日志: 已尝试全部无效: 补充:CLI 版 |
同款硬件复现成功(0.2.0-rc.2 win-x64),补充两个证据点环境
证据 1:主进程退出码与楼主日志签名一致PowerShell 与楼主贴的 证据 2:叠加发现——单实例锁冲突时应用会"零反馈静默退出"(另一个独立问题)当机器上同时运行着旧版桌面端(0.2.0-rc.1,共用用户数据目录 即旧实例持有 建议(在楼主 4 条基础上追加)
|
|
The fact that the process exits almost immediately and that the diagnostic information only goes to stdout makes the first problem observability rather than the exact GPU failure. I’d try launching the executable from a terminal instead of the normal shortcut so the Electron/Chromium stderr/stdout output remains visible. I’d also test whether disabling GPU acceleration changes the behavior. If the application starts with GPU acceleration disabled, that would strongly narrow the issue to Chromium/Electron GPU initialization rather than the application startup sequence itself. For example, depending on how the packaged application exposes Chromium arguments, testing with a GPU-disabled launch option can help isolate the problem. Because the environment includes Intel Arc and multiple virtual displays, I’d also compare: Intel Arc + virtual displays + GPU acceleration The goal is to identify whether the failure follows the GPU, the virtual-display configuration, or Electron’s GPU process initialization. For the application itself, writing startup diagnostics to a persistent log file would also make this class of failure much easier for users to diagnose because a normal desktop launch doesn’t expose stdout. If this clears up the issue, please consider marking the answer as accepted so others who run into the same problem can find the solution more easily. |
Uh oh!
There was an error while loading. Please reload this page.
问题概述
Windows 11 桌面端无法通过应用自己的快捷方式启动,双击后进程约 1 秒内退出,无任何界面、无任何提示。稳定复现。
同时发现:应用不写任何日志文件,所有诊断信息只走 stdout,用户双击启动时完全看不到,导致无法自助排查。
环境信息
硬件
显示适配器(关键)
PCI\VEN_8086&DEV_7D55&SUBSYS_3E7219E5&REV_08ROOT\DISPLAY\0000ROOT\DISPLAY\0001系统中并存 1 个物理 GPU + 2 个虚拟显示器适配器,怀疑这是 GPU 枚举失败的直接诱因。
软件版本
deepseek-harness-0.2.0-rc.1-win-x64.exe(288,472,536 字节)deepseek-harness-0.2.0-rc.2-win-x64.exe(289,313,640 字节)resources/version=44.0.0)D:\software\DeepSeek Harness%APPDATA%\ZCode复现步骤
DeepSeek Harness.exe(即不带任何命令行参数 —— 这是普通用户的标准路径)。要看日志需要手动重定向 stdout:
实际表现
崩溃日志
exit_code=-2147483645即0x80000003(STATUS_BREAKPOINT)。GPU 进程连续崩溃约 10 次后,Chromium 判定 GPU 不可用,直接终止整个应用,连主界面都进不去。命令行参数兼容性实测
--disable-gpu--disable-gpu --disable-gpu-compositing--disable-gpu-compositing--use-angle=swiftshader--disable-gpu-sandbox--in-process-gpu--no-sandbox目前唯一可行的规避手段是
--no-sandbox。但这会关闭 Chromium 渲染沙箱,不适合作为面向普通用户的默认方案 —— 用户不知道要加这个参数,只会认为「软件坏了」。附带问题:应用不写日志文件
排查时发现
%APPDATA%\ZCode下只有 Chromium 会话数据(Cache/GPUCache/Local Storage/Network等),没有任何logs/目录或.log文件。全部诊断输出只走 stdout,用户双击启动时被丢弃,导致:用户看不到崩溃原因、无法通过提交日志报 bug、只能靠人工重定向 stdout 才能定位问题。
建议增加本地日志落盘(如
%APPDATA%\ZCode\logs\main.log,带滚动),对桌面端问题定位帮助极大。期望表现
补充
0.2.0-rc.1与0.2.0-rc.2的崩溃行为完全一致,仅堆栈行号不同(main.js:11565→main.js:12067),说明这不是某个版本的回归,而是持续存在的兼容性问题。ROOT\DISPLAY设备后再尝试复现。All reactions