Runtime and Models

Runtime and Models / Runtime 与模型

← Wiki home / 手册首页 · English · 中文 · Troubleshooting / 故障排查

Use this page when / 适用情况： Runtime is detecting, missing, or needs repair; a model is unavailable; or you need to understand CPU, CUDA, Auto, Windows Commit, RAM, and VRAM.
Runtime 正在检测、缺失或待修复，模型不可用，或者需要理解 CPU、CUDA、Auto、Windows Commit、内存和显存。

English

Why Runtime is separate

The Windows application package stays relatively small by keeping heavyweight local-AI libraries outside the EXE tree. The active data root contains a shared Runtime sidecar that healthy later versions can reuse.

Runtime provides executable libraries such as PyTorch, sentence-transformers, LanceDB, Arrow, and related dependencies. Model files are managed separately. A healthy Runtime does not automatically mean that every required model is present, and downloaded CUDA files do not prove that the current machine can use CUDA.

CPU, CUDA, and Auto

• CPU works without an NVIDIA GPU and is the most compatible option, but semantic work is slower.
• CUDA requires a compatible NVIDIA GPU, a working driver, the CUDA Runtime profile, enough GPU memory, and enough Windows Commit/physical memory.
• Auto is recommended. It uses CUDA when the full chain is ready and the current workload fits, then falls back or skips optional reranking when necessary.

A system CUDA toolkit by itself is not enough. Caelune still needs its own matching Runtime payload.

Installation

Use the Runtime page in the desktop app. The packaged build uses its bundled installer-side Python and downloads a manifest-locked wheel set before validating imports.

The process is staged:

1. download
2. verify
3. install offline from the downloaded payload
4. validate modules and device availability

If the application is using the active Runtime during repair, the new payload may be staged as pending and activated at the next launch.

Models

The default semantic stack uses:

• BAAI/bge-m3 for embeddings
• BAAI/bge-reranker-v2-m3 for optional second-pass reranking

Model downloads need disk space in addition to Runtime and index space. Do not manually move only part of a model directory.

Windows Commit is not the same as free RAM

Windows Commit measures committed virtual memory against the combined RAM and paging-file limit. A computer can show free physical memory and free GPU memory while Commit headroom is still low.

Recommended actions when Commit is repeatedly constrained:

• keep the device setting on Auto
• close other large AI, browser, virtual-machine, or data-processing workloads
• enable “Automatically manage paging file size for all drives” in Windows, or choose an appropriate paging-file size
• restart Windows if a third-party process has not released committed memory

Low Commit does not normally require reinstalling Runtime or models.

Logs

Runtime installation and validation logs live below the active data root. Use the Runtime page to open the actual location instead of assuming a fixed AppData path.

Technical reference: RUNTIME_SETUP.md.

中文

为什么 Runtime 与 EXE 分离

Windows 发布包把大型本地 AI 依赖放在 EXE 目录之外，以控制软件包体积。当前数据目录中的共享 Runtime 可以被后续版本复用，不需要每次更新软件都重新下载整套组件。

Runtime 提供 PyTorch、sentence-transformers、LanceDB、Arrow 等执行库；模型文件单独管理。因此“Runtime 健康”不代表所有模型一定完整，下载了 CUDA 文件也不代表当前电脑一定能够使用 CUDA。

CPU、CUDA 与 Auto

• CPU 不要求 NVIDIA 显卡，兼容性最好，但语义处理速度较慢。
• CUDA 需要兼容的 NVIDIA 显卡、正常驱动、CUDA Runtime、足够显存，以及足够的 Windows Commit 和物理内存。
• Auto 是推荐选项。完整条件满足时使用 CUDA；当前任务资源不足时安全降级，或者只跳过可选重排。

仅仅在系统中安装 CUDA Toolkit 并不够，星野仍需要自己的匹配 Runtime。

安装流程

优先在软件的 Runtime 页面操作。发布包会使用自带的安装用 Python，下载由清单锁定的 wheel 文件，然后再进行模块验证。

流程分为：

1. 下载
2. 校验
3. 使用已下载文件离线安装
4. 验证模块和设备状态

如果当前程序正在使用旧 Runtime，修复内容可能先进入待应用目录，在下一次启动时启用。

模型

默认语义模型为：

• BAAI/bge-m3：生成向量
• BAAI/bge-reranker-v2-m3：可选的二次重排

模型除了 Runtime 和索引之外还需要额外磁盘空间。不要只手动移动模型目录的一部分。

Windows Commit 不等于空闲物理内存

Windows Commit 是已经承诺的虚拟内存与“物理内存 + 分页文件上限”的关系。因此即使任务管理器仍显示空闲内存和显存，Commit 余量也可能不足。

如果经常出现 Commit 受限：

• 保持设备为 Auto
• 关闭其他大型 AI、浏览器、虚拟机或数据处理程序
• 在 Windows 中启用“自动管理所有驱动器的分页文件大小”，或者设置合适的分页文件
• 如果第三方进程没有释放提交，必要时重启 Windows

单纯 Commit 不足通常不需要重装 Runtime 或模型。

日志

Runtime 安装和验证日志保存在当前数据目录下。请通过 Runtime 页面打开实际位置，不要假定固定的 AppData 路径。

技术参考：RUNTIME_SETUP.md。

---

Previous / 上一篇： ← Getting Started / 快速开始 · Next / 下一篇： Knowledge Bases and Live Watch / 知识库与热监听 →