--- title: Strata-DS_Windows canonical_url: "https://www.modelscope.cn/models/windecay/Strata-DS_Windows" md_url: "https://www.modelscope.cn/models/windecay/Strata-DS_Windows.md" repository: windecay/Strata-DS_Windows chinese_name: Strata-DS_Windows last_updated: 2026-10-29 pipeline_tag: visual-question-answering tasks: - visual-question-answering downloads: 0 stars: 3 --- # Strata-DS_Windows > Strata-DS_Windows - windecay 在 ModelScope 开源的模型。Strata DS Windows 轻量独立包含启动器 by ✿冰華 - **Repository**: windecay/Strata-DS_Windows - **Tasks**: visual-question-answering - **Downloads**: 0 - **Stars**: 3 - **Last updated**: 2026-10-29 Source: https://www.modelscope.cn/models/windecay/Strata-DS_Windows --- # Strata DS Windows 1.2.5 by ✿冰華 本地运行 DeepSeek-V4.1-Flash-1017，包含 Hyoka 桌面启动器、App 工作区、中英文聊天网页、Python、Qt 和 NVIDIA 引擎。启动器版本 **1.2.5**，上游 Strata **0.2.7**，引擎及 Control **0.1.52**。无需另装 Python、Node、CUDA Toolkit 或编译工具。 Run DeepSeek-V4.1-Flash-1017 locally with the Hyoka desktop launcher, App workspace, bilingual Web chat, Python, Qt and NVIDIA engine included. Launcher **1.2.5**, upstream Strata **0.2.7**, engine and Control **0.1.52**. No separate Python, Node, CUDA Toolkit or compiler installation is needed. ## 启动 1. 双击 `Start-Strata-DS.bat`。打开窗口不会自动加载模型。 2. 首次使用进入“模型管理”，选择魔搭、Hugging Face 或 hf-mirror、量化档位和下载目录。已有文件可选择原目录，点击“生成配置”，不用重新下载主模型。 3. 选择模型及“标准模式”或“Heretic 模式”，点击“启动服务”。Heretic 启用随包 LoRA，DSpark 是独立的草稿开关。 4. 服务就绪后点击“打开 App 工作区”或“打开聊天页”，也可以复制 API 地址连接其他程序。默认仅监听本机。 5. “停止服务”结束本窗口创建的服务。其他窗口启动的服务请在原窗口停止。 Double-click `Start-Strata-DS.bat`. Opening the launcher does not load a model. For a new installation, use **Model manager** to download and prepare the selected components. For existing weights, select their folder and use **Prepare profile**. Select Standard or Heretic mode, then **Start service**. Heretic uses the bundled LoRA; DSpark drafting is independent. Once ready, open the App workspace or Chat, or copy the API address. The default address is local only. Stop a service from the launcher that started it. ## 1.2.5 更新 - 在“启动参数 → 加载 → 多磁盘模型加载”增加可选的主干分盘配置，默认关闭。四个主干分片可分别指定路径，不限于 2+2，也可分布到多块 SSD。2+2 按钮将第 1、3 个保留原位置，第 2、4 个指定到第二目录。 - 支持直接选择已有分片；也可以勾选复制缺失分片，显示复制和校验进度并支持取消。原文件保留，完整 SHA-256 校验及引擎文件检查通过后才允许启用新布局。 - 显示分片所在的物理磁盘；四个分片都位于同一物理磁盘时会提示共享带宽。Engram、DSpark、视觉投影、LoRA 和模型参数保持原有设置。 - 准备文件不改变当前配置。停止服务后点击“启用已准备的配置”，下次启动生效；“恢复原文件布局”可以返回原路径，同时保留后来调整的模型参数。 Version 1.2.5 adds optional **Settings → Loading → Multi-disk model loading**, off by default. Each of the four main shards can use a separately selected path, including layouts across more than two SSDs. The 2+2 shortcut keeps shards 1 and 3 at their original locations and assigns shards 2 and 4 to a second folder. Use existing files or copy missing shards with progress and cancellation. Original files are kept. SHA-256 verification and the engine file check must pass before activation. The panel identifies physical disks and warns when all four shards share one physical disk. Engram, DSpark, vision, LoRA and model settings retain their existing configuration. Stop the service before enabling a prepared layout or restoring the original locations. ### 多磁盘使用步骤 1. 选择已准备的模型，保存尚未保存的参数，进入“启动参数 → 加载”。 2. 在“多磁盘模型加载”中逐行输入各分片的完整文件路径，或点击 `…` 选择已有文件。也可以选择第二个模型目录，自动填入 2+2 路径后再按需修改；重新选择目录或点击“设置 2+2 路径”会重新填写四行路径。 3. 目标位置缺少分片时，填写包括原始文件名的完整目标路径，并勾选“将缺少的分片复制到所选路径，并保留原文件”。点击“校验并准备配置”，等待校验完成。文件名和内容需要与所选模型的对应分片一致。 4. 停止模型服务，点击“启用已准备的配置”，然后重新启动服务。要取消分盘，停止服务后点击“恢复原文件布局”。 需要两块独立 SSD 才可能获得叠加读取收益；同一 SSD 的两个盘符共享带宽。启动加速取决于磁盘、模型、缓存容量及系统文件缓存，不代表生成速度提升。磁盘断开或文件损坏会阻止使用该布局；连接原磁盘或恢复文件后重新校验。校验及复制可能读取大量数据，可在任务完成后再开始其他磁盘密集工作。 Select a prepared model, save pending parameter edits, and open **Settings → Loading**. Enter a complete file path or browse for each main shard. Alternatively, choose a second folder to fill the 2+2 paths, then change individual paths as needed. Choosing the folder again or selecting **Set 2+2 paths** replaces all four path fields. For missing files, enter the full destination paths with the original filenames and enable copying. Filenames and contents must match the corresponding shards of the selected model. Choose **Verify and prepare layout**, wait for validation, stop the service and choose **Enable prepared layout**. To return, stop the service and use **Restore original layout**. Independent SSDs may reduce startup time; two partitions on one SSD share bandwidth. Results depend on storage and cache settings and do not establish faster generation. Reconnect an unavailable disk or restore missing files before retrying. ### 惩罚参数 1.2.5 升级包包含支持重复、存在和频率惩罚的引擎。DeepSeek 默认设置不启用这些惩罚；未设置惩罚时，采样使用无惩罚路径。 在“启动参数 → 生成 → 默认生成参数”中设置“重复惩罚”“存在惩罚”“频率惩罚”和“惩罚统计范围”，也可通过 API 请求传入对应参数。升级不会修改已有参数，但之前保存的非默认值、聊天设置或客户端发送的惩罚可能开始生效，生成结果可能改变。启用重复思考恢复时，恢复过程也可能使用存在惩罚。 设置默认参数时，存在惩罚和频率惩罚请使用非负数，重复惩罚必须大于 0。使用惩罚但未填写统计范围，或将统计范围设为 0 时，服务使用最近 64 个 token。 ### Sampling Penalties The 1.2.5 update includes an engine that handles repetition, presence and frequency penalties. The default DeepSeek settings do not enable them; sampling uses the no-penalty path when no penalty is set. Set **Repetition penalty**, **Presence penalty**, **Frequency penalty** and **Penalty history window** under **Settings → Generation → Default generation**, or pass the corresponding fields in an API request. The update does not change settings, but previously saved non-default values, Chat settings or client-supplied penalties may begin to affect output. When reasoning-loop recovery is enabled, its recovery process may also use a presence penalty. For saved defaults, presence and frequency penalties must be non-negative and repetition penalty must be greater than zero. When a penalty is active and its window is unset or set to zero, the service uses the last 64 tokens. ## 1.2.4 更新 - 升级包含 App 工作区及其运行组件。启动器始终显示“打开 App 工作区”；服务就绪后手动打开，组件不完整时会提示。 - App 工作区基于上游 0.2.6 的适配版本；模型服务和引擎继续使用上游 0.2.7 与 0.1.52。 - 显存休眠状态增加开启指引和“休眠设置”入口。显存休眠仍默认关闭，保存设置并重启服务后生效。 - 保留原模型、个人配置、聊天数据和缓存；本次更新不自动下载模型。 Version 1.2.4 includes the adapted App workspace based on upstream 0.2.6; the model service and engine remain on upstream 0.2.7 and engine 0.1.52. The launcher always shows **Open App workspace**; open it after the service is ready. If the component is incomplete, the launcher explains why it cannot open. GPU sleep now has a settings shortcut and activation guidance; it remains off by default. Existing models, settings, conversations and caches are preserved. ## 1.2.3 更新 源码已合入上游 0.2.7，Control／引擎为 0.1.52。包含 DeepSeek 提示词读取与缓存更新，以及网页会话列表、编辑与回退、排队发送、公式渲染和 HTML/SVG 预览。本地双语页面、LoRA、DSpark、视觉及显存恢复功能保留。会话列表、编辑和公式渲染可在聊天设置中启用。 上游增加 512K、1M 上下文选项，仍受模型文件上限约束。已保存的参数继续生效。需要比较旧默认行为时，可在启动环境设置 `DS4_DEFAULTS=026`；明确设置的参数仍优先。更新只替换清单中的程序文件，原模型与个人配置保留。 Version 1.2.3 includes upstream 0.2.7 with Control/engine 0.1.52: DeepSeek prompt/cache updates and Web chat history, editing/rewind, queued messages, math and sandboxed HTML/SVG previews. Local bilingual pages, LoRA, DSpark, vision and GPU recovery remain available. Enable saved chats, editing and math in Chat settings. The 512K/1M context choices remain subject to the model's limit. Saved settings remain in effect. `DS4_DEFAULTS=026` selects the upstream 0.2.6 default rules, while explicit settings take priority. The update replaces only listed program files and preserves models and personal settings. ## 1.2.2 更新 - 模型管理支持选择 DeepSeek 主模型的“原始版本”和“更新版本”。更新版本使用上游修正聊天模板及图片格式后的首分片；其余分片和共享组件可以复用。两版分别生成启动配置，原始版本仍是默认选择。 - 修复魔搭文件更新后与原固定清单不一致导致的下载报错；现在按所选版本使用对应的固定文件信息。新增 hf-mirror 下载源，可手动切换下载源并继续下载相同文件。 - “视觉投影”支持 Unsloth 和 ISTA-DASLab 的 BF16 文件，可与两版主模型分别搭配。默认仍使用 Unsloth；两种文件会在生成配置前校验 SHA-256。 - 程序更新不会自动下载、改写或替换已有模型及配置。本机没有更新版首分片时需要单独下载，其余已有文件继续复用。 Version 1.2.2 lets you choose the original or updated DeepSeek main model. The updated first shard contains the upstream chat-template and image-format corrections; the other shards and shared components can be reused. Each version has its own launch profile, and the original remains the default. ModelScope now uses the pinned files for the selected version, and hf-mirror is available as a download source. Choose the Unsloth or ISTA-DASLab BF16 vision projector independently; Unsloth remains the default and each file is verified before profile preparation. Updating the program does not download or replace existing models and settings. The updated first shard is a separate download if you do not already have it. ## 1.2.1 更新 - 增加 **59 项高级设置**，分为生成、性能、加载、兼容、多卡和 MCP。可调整思考与采样、会话缓存、专家分配、预填充、按需加载、自动卸载、超时和客户端兼容选项。没有指定的值继续使用原生默认，保存其他参数不会改写它们。 - “性能”页显示专家热区文件、更新时间和累计计数，支持更改路径、刷新及打开目录。热区在请求结束及正常停止时保存，下次启动继续读取。它与会话缓存分别管理。 - 增加 DeepSeek **视觉识图**组件和开关，支持聊天网页、OpenAI 与 Anthropic 图片请求。未准备视觉组件时明确显示仅支持文本，运行状态显示实际是否启用视觉。 - 优化上下文、缓存、CPU 线程和草稿长度的显示，并显示本机硬件。专家 RAM 缓存区分自动、关闭和指定容量，原参数继续保留。 - 日志刷新保留阅读位置、选中文本和横向滚动。向上阅读暂停跟随，勾选“跟随最新”可返回末尾。界面显示最近 1,800 行，完整日志仍保存在文件中。 - 合入上游 0.2.6，包括专家自动预载和 Windows 内存提交额度检查。是否自动预载取决于参数和实际内存条件。 Version 1.2.1 adds **59 advanced settings** under Generation, Performance, Loading, Compatibility, GPUs and MCP. Unset options retain native defaults, and unrelated saves preserve existing choices. The Performance page shows the expert heat file, modification time and counters. Heat is saved after requests and normal shutdown and reused at the next start. DeepSeek vision now works through Web Chat, OpenAI and Anthropic image requests. Hardware and parameter displays are clearer, and log updates retain the reader's position and selection. Upstream 0.2.6 brings expert preloading and Windows commit-budget checks; automatic preloading remains conditional. ## 视觉识图 在“模型管理”保留所需的主模型、Engram、运行元数据及 DSpark、Heretic 选项，额外勾选“视觉权重”，并在“视觉投影”选择 Unsloth 或 ISTA-DASLab，下载或复用对应文件后生成配置。视觉文件约 **0.91 GiB**，支持魔搭、Hugging Face、hf-mirror 和完整 SHA-256 校验；发布包不包含这些权重文件。 准备完成后，在“启动参数 → 常规 → 模型功能”启用“视觉识图”，保存并重启服务。聊天页可上传图片。已有主模型无需搬动或重新下载。未准备视觉的旧配置继续使用文本模式，升级程序不会自动下载模型或改写用户配置。 In Model manager, keep the components you use and also select **Vision weights**. Choose the Unsloth or ISTA-DASLab projector, download or reuse the file, then prepare the profile. Each projector is about **0.91 GiB**, is available from ModelScope, Hugging Face and hf-mirror, and is verified with SHA-256. Projector weights are not included in the program archive. Enable **Read pictures** under **Settings → General → Model features**, save and restart the service. Existing text-only profiles remain text only until vision is prepared. ## 缓存与模型卸载 “性能”页可设置提示词复用、内存会话数量、最短上下文、磁盘缓存预算、保留期限及保存目录。**持久会话缓存目前仅保存纯文本会话，不保存含图片的会话。** 专家 RAM 缓存保存的是模型权重，与会话缓存不同。 网页右上角可切换中英文并记住选择。聊天页顶部提供“加载模型”和“卸载模型”：卸载释放显存及模型 RAM 缓存，Web/API 服务、聊天记录和草稿继续保留；下一次请求会重新加载模型。正在生成或有请求排队时不能卸载。 The Performance page controls prompt reuse and conversation caching. **Persistent conversation caching currently covers text only; picture conversations are not stored.** The expert RAM cache holds model weights and is separate. The Web language selector remembers your choice. **Unload model** releases GPU memory and the model's RAM cache while retaining the Web/API service, chat history and draft. The next request reloads the model. Unloading is blocked while requests are running or queued. ## 显存休眠（测试中） 1. 在“启动参数 → 加载 → 模型加载与卸载”开启“启用显存休眠”，保存设置并重启服务。此功能默认关闭，仅支持配置为使用单张 NVIDIA GPU 的服务。 2. 等待请求结束，点击“显存休眠”。模型的主要显存数据会保存到内存；模型进程和已有上下文保留。未开启自动唤醒时，休眠或状态切换期间的生成请求会被拒绝。 3. 需要继续使用时，先释放足够的显存，再点击“唤醒显存”，或使用下面的请求自动唤醒。唤醒失败时数据仍保留在内存中，可腾出资源后重试。手动卸载模型或停止服务会丢弃这份休眠数据。 同一设置页还提供两项自动控制，保存后重启服务生效，默认均关闭： - **空闲后自动休眠显存**（`idle_gpu_sleep_s`）：单位为秒，0 关闭。例如 300 表示请求全部结束、空闲 5 分钟后休眠。启用且引擎支持时，它优先于“空闲后自动卸载”，继续保留 RAM。 - **收到推理请求时唤醒显存**（`gpu_sleep_auto_wake`）：唤醒后继续原请求，支持普通与流式响应。并发请求共用一次唤醒；若失败，本批请求收到错误，内存状态保留，后续请求可以重试。查询状态不会唤醒。 休眠需要额外的空闲内存保存显存数据，并在检查时预留至少 8 GiB。已有模型 RAM 缓存不会因此释放；CUDA 运行环境和部分临时显存也会保留，不能按完全清空显存估算。空闲内存不足、模型正在处理请求或后台工作尚未结束时，休眠会被拒绝。 单卡完整 DeepSeek 模型的重复休眠／唤醒、后续文本与图片生成、上下文复用和休眠后卸载已通过功能验证，包括标准与 Heretic 模式。测试为显存数据副本预留了内存；较大的专家 RAM 缓存可能需要调低。其他硬件、长时间压力测试及真实显存耗尽后的恢复仍待验证，此功能继续默认关闭。 GPU sleep is experimental and off by default. Under **Settings → Loading → Model lifecycle**, enable **GPU sleep**, save, and restart the service configured for one NVIDIA GPU. When requests have finished, use **Sleep GPU**; use **Wake GPU** manually or enable automatic wake. Without automatic wake, requests are rejected while the GPU is asleep or changing state. Waking needs enough free VRAM; on failure the saved state remains in RAM for another attempt. Explicitly unloading the model or stopping the service discards it. The same page adds two opt-in controls, applied after saving and restarting. **Sleep GPU after inactivity** (`idle_gpu_sleep_s`) sets the idle seconds; 0 disables it and 300 means five minutes after requests finish. When supported it takes priority over idle full unloading and retains RAM. **Wake GPU on inference request** (`gpu_sleep_auto_wake`) wakes and continues ordinary or streamed requests. Concurrent callers share one wake attempt; failure preserves the saved state and allows a later retry. Status queries never wake the model. Sleep needs additional available RAM for the saved GPU data and checks for an 8 GiB reserve. Existing model RAM caches remain allocated. CUDA and some temporary GPU allocations also remain, so sleep does not release all VRAM. Repeated full-model sleep/wake, subsequent text/image generation, context reuse and unloading while asleep passed on a single NVIDIA GPU, covering Standard and Heretic modes. The tests reserved RAM for the GPU snapshot; larger expert-cache settings may need reducing. Other hardware, prolonged stress and recovery from actual VRAM exhaustion remain unvalidated. The feature stays off by default. ## App 工作区 本次升级包包含 App 工作区组件。启动器始终显示“打开 App 工作区”，组件缺失或不完整时会提示安装包含该组件的版本。 This update includes the App workspace. The launcher always shows **Open App workspace** and reports a missing or incomplete component. 启动器的“打开 App 工作区”使用本机浏览器，自动连接本窗口启动的模型服务，并读取实际上下文长度和视觉状态。原聊天页与 API 地址继续保留。停止服务或关闭启动器时，App 和它启动的 Agent 一同停止。 App 的 Agent 会在需要时压缩较早的对话，也可以在对话中输入 `/compact` 手动压缩。压缩会用摘要代替模型上下文中的部分历史，可能遗漏细节；单条输入超过上下文容量时仍需缩短输入。原聊天页不因此获得同样的压缩功能。 App 对话保存在 `userdata/app/`，工作区文件保存在 `workspaces/`。更新或移动程序时保留这两个目录。App 只监听本机，使用启动器当前选择的语言打开。 Open **App workspace** after starting the service. The browser App connects to the model owned by this launcher and uses its reported context and vision capability; the original Chat page and API remain available. The App and its Agent stop with the service. The Agent can summarize older context automatically or with `/compact`. Summaries can omit details, and a single oversized input still needs shortening. App chats are stored in `userdata/app/` and workspace files in `workspaces/`; keep both when upgrading. ## 目录与升级 | 目录 | 内容 | | --- | --- | | `app/` | 启动器、引擎、Python、Qt 和服务程序 | | `models/` | 已下载模型与适配器 | | `config/` | 个人设置及启动配置 | | `cache/` | 生成的 pack、专家热区与会话缓存 | | `logs/` | 运行日志 | | `userdata/app/` | App 对话与设置 | | `workspaces/` | App 工作区文件 | 移动整个文件夹时保留内部结构；用户选择的外部模型目录需要继续可访问。完整包仅包含程序和随包 Heretic LoRA，不包含主模型、视觉权重、个人设置、API key、日志、App 对话或缓存。 使用完整包升级时，关闭旧启动器和服务，将程序文件覆盖到原目录，**保留 `models/`、`config/`、`cache/`、`logs/`、`userdata/` 和 `workspaces/`**。使用 1.2.4 → 1.2.5 差异更新包时，将更新包完整解压到任意目录，双击 `Update-Strata-DS.bat`，在弹出的窗口选择原 Strata DS 安装目录；也可以在终端指定原安装路径。更新程序会校验适用版本及文件、备份替换内容，并保留个人目录。README 可以自行修改或删除：已有文本会先按原字节备份，再替换为新版；其他程序文件继续严格校验。取消选择不会修改文件，检查失败时不会继续覆盖。 Keep the folder structure when moving the installation. External model folders must remain accessible. The full archive includes the program and bundled Heretic LoRA, excluding base models, vision weights, personal settings, keys, logs and caches. For an upgrade, close the old launcher and service, replace program files and **preserve `models/`, `config/`, `cache/`, `logs/`, `userdata/` and `workspaces/`**. For the 1.2.4 → 1.2.5 difference package, extract it anywhere, double-click `Update-Strata-DS.bat`, and select the existing Strata DS installation folder. Alternatively, pass the installation path in a terminal. You may edit or remove README: existing notes are backed up byte for byte before replacement. Other program files remain strictly checked. The updater preserves user folders. Cancelling the folder picker changes no files. ## 硬件与验证范围 Windows 64 位，NVIDIA RTX 20/30/40/50，需兼容 CUDA 13 的 NVIDIA 驱动。引擎含 `sm_75`、`sm_86`、`sm_89`、`sm_120` 架构代码。架构覆盖不代表每款显卡均已完成硬件验收。 此前的引擎已通过标准与 Heretic 模式识图、Heretic 图片文字读取、网页图片上传及卸载后重载验证。两模式的专家热区均通过请求后保存、重启继续累计检查。 Windows 64-bit with an NVIDIA driver compatible with CUDA 13. The engine includes `sm_75`, `sm_86`, `sm_89` and `sm_120` code. Architecture coverage does not establish hardware acceptance for every listed card. Image requests, OCR, browser uploads, reload and expert-heat persistence passed functional checks on an earlier engine. ## 显存被其他程序占用时 启用显存休眠后，启动器会显示恢复所需、当前可用和还缺多少显存。“恢复所需”包括尚未恢复的数据和额外余量，默认余量为 512 MiB。Windows 下同时检查 CUDA 与整张显卡的空闲显存，采用较小值。 | 当前情况 | 可以怎么做 | | --- | --- | | 想保留内存与当前上下文 | 点击“唤醒显存”。容量不足时保留 RAM 并等待；观察到资源连续至少 2 秒足够后自动恢复。 | | 其他 GPU 任务仍需要显卡 | 点击“取消等待”。释放显存后也不会自动唤醒，后续生成请求会提示已暂停；等其他任务结束后再点击“唤醒显存”。 | | 无法腾出原来那么多显存 | 点击“按可用显存重载”并确认。服务短暂断开，使用已保存设置重新安排 GPU 专家缓存；RAM 和上下文缓存会重建，聊天记录保留，生成可能变慢。 | | 希望以后少占显存 | 点击“调整显存占用”，停止服务，提高“预留显存”，保存后启动。它减少可用于 GPU 缓存的预算；“唤醒后保留的空闲显存”只影响恢复门槛，不会缩小原快照。 | 显存不足时只检查容量，不反复申请大块显存。实际分配失败后默认等待 15 秒再试；连续 3 次失败会暂停自动恢复，并显示处理建议。已经开始的休眠或唤醒传输不能直接取消。部分恢复若仍占有显存，状态会显示保留量；需要立即释放时可停止服务，但会丢弃 RAM 快照。 保留 RAM 的唤醒允许重新申请同样容量的物理显存，不能直接缩小原快照。按可用显存重载会丢弃旧快照；如果固定权重或必需缓冲区仍放不下，仍可能加载失败。这时请停止其他 GPU 任务腾出显存后重新启动；必要时调低上下文、关闭 DSpark／视觉等可选组件，或选用硬件能容纳的模型，并查看运行日志。驱动异常或引擎退出时，需要重新加载，不能继续保证旧 RAM 快照可恢复。 聊天页收到资源不足错误后，该请求尚未执行。等待模型恢复后点击“重试此消息”，不会重复添加用户消息，也不会清空后来输入的草稿。刷新页面后若附件内容已不可用，请重新附加图片或文件。其他程序可能继续申请显存，以上检查不能预留整张显卡的资源；需要优先保障其他 GPU 任务时请取消等待。 ### When another application is using VRAM With GPU sleep enabled, the launcher shows required, available and missing VRAM. Required memory includes the allocations still to restore plus headroom, which defaults to 512 MiB. On Windows the engine checks both CUDA and device-wide free memory and uses the smaller value. - **Wake GPU** keeps RAM and the current context. If capacity is insufficient, it waits without repeatedly allocating large GPU blocks. Automatic recovery starts after enough memory has been observed for at least two seconds. - **Cancel wait** pauses automatic recovery, including wake on subsequent requests. It stays asleep even after memory becomes free. Select **Wake GPU** when the other task has finished. A sleep/wake transfer already in progress cannot be cancelled directly. - **Reload to fit free VRAM** requires confirmation. It briefly disconnects the service and rebuilds the GPU expert cache using saved settings and currently available memory. RAM/context caches are rebuilt; chat history remains. Generation may be slower. This discards the previous snapshot even if loading fails. - **VRAM settings** opens the relevant settings. Stop the service, increase **Reserved VRAM**, save and start to reduce future GPU cache usage. **Free VRAM to keep after waking** changes the restoration threshold only; it cannot shrink a saved snapshot. An actual failed allocation waits 15 seconds before retrying by default. Three consecutive failures pause automatic recovery and show next steps. Partial recovery may retain some GPU allocations; the status shows the amount. Stopping the service releases them and discards the RAM snapshot. Restoration can use different physical GPU blocks of the same capacity, but cannot resize the saved state in place. If fixed weights or required buffers still do not fit after reloading, free VRAM from other tasks and start again. If necessary, lower context, disable optional DSpark/vision components or select a model that fits, and inspect Logs. Driver failures or an exited engine require a reload; the old RAM state may no longer be recoverable. A request rejected for insufficient VRAM has not executed. Once the model is ready, use **Retry this message** in Chat. This preserves newer drafts without adding the same user message twice. Reattach any images/files whose content was lost after refreshing the page. Capacity checks do not reserve the GPU against other applications; cancel waiting when their work should take priority.