Strata Manager v0.2.7

中文

• 无头 Server 实验版（#6 P0/P1）：新增 strata-manager-server（axum），默认 127.0.0.1:8787，首次运行生成 Bearer 令牌；POST /api/v1/rpc 对接 RuntimeState::dispatch，可托管同一套 Web 前端。浏览器侧增加 server 传输与令牌门禁；文件选择改为服务器路径输入；隐藏托盘、悬浮监控与桌面自动更新器。附 systemd unit、scripts/package-server.sh 与 docs/SERVER.md。CI 增加 server crate 测试与 Linux 打包产物；发布资产在原 11 项之外增加 Strata.Manager.Server_<ver>_amd64.tar.gz / .deb（共 13 项，仍无 SHA256SUMS，Server 包不进 updater）。
• macOS Metal 引擎预制包（#53 第二期 a）：在 Apple Silicon 上可下载并校验上游 strata-engine-*-macos-arm64-metal.tar.gz（钉死 v0.2.0-beta 资产名与 SHA256，应用内校验，不附带 SHA256SUMS），解包到 engine/；也支持选用本地 .tar.gz 或已有 Strata/engine、build-metal 目录。Mac 上不提供 Windows/Linux 引擎包；模型目录隐藏上游未在 Metal 验收的 Qwen3.6 / Swift / Coder；默认 --spec 2，不开放 PC --calibrate 与 --parallel。./setup.sh 源码构建仍为手动提示，真机推理验收待进行。
• 任务栏监控（Windows，#71）：「监控设置 → 任务栏监控（Windows）」新增开关，默认关闭。开启后在主屏任务栏通知区左侧嵌入一个小部件：圆环显示 GPU 负载（可改为显存占用或不显示），旁边每项指标一列：上方是加粗数值和小号单位（如「21.3 t/s」），下方是灰色短标签（生成、预填充、GPU、显存、温度）。每项指标有自己的颜色（生成绿、预填充蓝、GPU 橙、显存紫），GPU 温度按数值变色（低于 60℃ 绿、60–80℃ 琥珀、高于 80℃ 红）；深浅两套配色在 Win11 浅色和深色任务栏上的对比度都不低于 4.5:1。圆环用所选指标的颜色，右下角小圆点表示服务状态。数字紧凑显示（3.2k，100 以上不带小数），每列宽度固定，数值变化时不会左右跳动；默认内容在 100% 缩放下宽 134 px。「颜色」区可为生成、预填充、GPU、显存、温度分别设置自定义颜色（色块 + 8 个在深浅任务栏上都清晰的预设色 + 自定义取色，可一键恢复默认）；自定义颜色在深浅两种任务栏上都使用，对比度不足时给出提示，自定义温度颜色会取代按温度变色；服务状态圆点颜色不可改。任务栏右键菜单新增「自定义颜色…」，直接打开设置页的颜色区。还可加入 GPU 负载、显存、GPU 温度、服务状态，宽度随所选内容变化。小部件由独立的原生 Win32 辅助进程绘制（不使用 WebView），每秒最多刷新一次，数值不变时不重绘；文字颜色跟随任务栏深浅色，按 DPI 缩放；任务栏自动隐藏、全屏应用、通知区变宽和 Explorer 重启时会跟随或重新挂载，放不下时回到桌面浮条并说明原因。左键打开 Strata Manager，不抢焦点、不出现在 Alt-Tab；右键菜单可「打开 Strata Manager」、在「自定义内容」中勾选显示项和圆环指标，或「关闭任务栏监控」，与设置页同步。仅支持横向的主屏任务栏。
• 托盘图标恢复为普通应用图标：撤回此前在托盘图标上绘制读数的实验方案，实时读数改由任务栏监控提供。

English

• Headless Server experiment (#6 P0/P1): adds strata-manager-server (axum), default bind 127.0.0.1:8787, first-run Bearer token; POST /api/v1/rpc maps onto RuntimeState::dispatch and can serve the existing Web UI. Browser gains a server transport and token gate; file picks become server path inputs; tray, floating monitor and desktop updater are hidden. Includes systemd unit, scripts/package-server.sh and docs/SERVER.md. CI tests and packages the server on Linux; releases add Strata.Manager.Server_<ver>_amd64.tar.gz / .deb (13 assets total, still no SHA256SUMS; server packages are updater-independent).
• macOS Metal prebuilt engine (#53 phase 2a): on Apple Silicon, Manager can download and verify the upstream strata-engine-*-macos-arm64-metal.tar.gz (pinned v0.2.0-beta asset name and SHA256; in-app checks only, no SHA256SUMS files), extract into engine/, or use a local .tar.gz / existing Strata engine or build-metal directory. Windows/Linux engine packages are never offered on Mac; the catalog hides Qwen3.6 / Swift / Coder (not measured on Metal upstream); defaults use --spec 2 and omit PC --calibrate / --parallel. ./setup.sh remains a manual hint; real-device inference validation is still pending.
• Taskbar monitor (Windows, #71): Monitor settings → "Taskbar monitor (Windows)" adds an on/off switch, off by default. When on, a small widget sits left of the notification area on the primary taskbar: a ring shows GPU load (or VRAM usage, or no ring), followed by one column per metric: a bold value with a small unit on top (e.g. "21.3 t/s") and a short grey label below (Gen, Prefill, GPU, VRAM, Temp). Each metric has its own colour (generation green, prefill blue, GPU orange, VRAM violet), and GPU temperature changes colour with the reading (green below 60 ℃, amber 60–80 ℃, red above 80 ℃); the light and dark palettes keep at least 4.5:1 contrast on Win11 light and dark taskbars. The ring uses the selected metric's colour, and a small dot at its lower right shows the service state. Numbers are compact (3.2k, no decimals above 100) and every column has a fixed width, so values never shift sideways; the default set is 134 px wide at 100 % scaling. A "Colors" section lets you override the colour of generation, prefill, GPU, VRAM and temperature (swatch, eight presets that stay readable on both light and dark taskbars, a custom picker, and reset to default); a custom colour is used on both taskbar themes, low contrast is flagged, and a custom temperature colour replaces the graded scheme, while the service-state dot keeps its fixed colours. The right-click menu adds "Customize colors…", which opens that section in the app. GPU load, VRAM, GPU temperature and service status can be added, and the width follows the selection. It is drawn by a separate native Win32 helper process (no WebView), refreshes at most once per second and skips repaints when nothing changed; text follows the taskbar's light/dark colour and DPI scaling; it follows or re-attaches on auto-hide, fullscreen apps, notification-area width changes and Explorer restarts, and falls back to the desktop strip with a reason when there is no room. Left-click opens Strata Manager without stealing focus or appearing in Alt-Tab; the right-click menu offers "Open Strata Manager", a "Customize" submenu for items and the ring metric, and "Close taskbar monitor", all in sync with the settings page. Horizontal primary taskbars only.
• Tray icon is the plain app icon again: the experimental readings drawn into the tray icon were withdrawn; live readings now come from the taskbar monitor.

Strata Manager v0.2.6

中文

• 通过「模型 - 添加目录」加入的 Strata 能正常启动（#78）：启动、依赖检测、一键修复、向导和 MCP 自检统一使用同一个解释器选择规则，和目录自己的 run-*.bat 一致（启动脚本指定的 Python、目录内 .venv、嵌入式 Python），并带上脚本设置的 PATH / PYTHONPATH，清掉系统的 PYTHONHOME；目录有自带环境时绝不改用系统 Python。配置里的引擎 exe 不存在时，扫描和启动前就会提示，不再等到加载模型时才报 WinError 2。加载失败不再只显示 error sending request for url (.../load)，而是按进程状态和引擎日志分成「引擎在加载模型时崩溃」「Python 环境不对」和仍在加载三种情况，原始错误保留在启动日志和诊断依据里；Windows 中文系统下的报错也不再乱码。
• 发布不再附带 SHA256SUMS：从下一个版本起，GitHub 发布页只提供安装包、对应的 .sig 更新签名和 latest.json，共 11 项资产。应用内更新仍按 latest.json 与 .sig 验签；应用内对模型和引擎 ZIP 的大小与 SHA256 校验不变。已发布版本的资产保持原样。
• 「使用已有模型」面板整理：向导与下载模型页的本地 GGUF 面板统一左对齐与间距节奏；路径放入输入框；量化行改为名称/说明在左、体积与状态点在右；状态用中性文字 + 小圆点，仅问题用色；MTP 与等效命令默认折叠；长说明收成一行。
• 使用已有模型（#75）：向导「选择模型」一步和「下载模型」页新增「使用已有模型」。选择存放 GGUF 的文件夹或其中一个分片后，Manager 按模型与量化版本分组，检查分片是否齐全、可读、GGUF 文件头和大小，再以官方安装器的 --gguf-dir 原地使用这些文件，不再下载模型权重；可选择已有的 MTP 草稿层（上游 setup.py 没有 --mtp 安装参数，因此映射为 --data-dir），并显示等效的 START-HERE.bat 命令。文件名与官方发布名不同但量化受支持时给出提示而不再拒绝；OrcaRouter 去审查 IQ3_XXS 走 ORCA.md 兼容流程；引擎不支持的量化会说明原因。已检测到但尚未登记的 Strata 运行环境现在也可直接用于准备模型，不再显示“请先选择 Strata 运行环境”。

English

• Strata folders added with Models → Add folder start correctly (#78): starting, the dependency check, one-click repair, the wizard and the MCP self-check now share one interpreter rule that matches the folder's own run-*.bat (the launcher's Python, the folder's .venv, or an embedded Python), keep the PATH / PYTHONPATH the script sets and drop the host PYTHONHOME; a folder with its own environment never falls back to the system Python. A configured engine exe that does not exist is reported at scan time and before starting, instead of as WinError 2 while loading. A failed load no longer shows just error sending request for url (.../load): it is classified from the process state and engine log as "engine crashed while loading", "wrong Python environment" or still loading, with the raw error kept in the startup log and diagnostic evidence; Windows error text on Chinese systems is no longer garbled.
• Releases no longer include SHA256SUMS: starting with the next version, GitHub releases provide the installers, their .sig updater signatures and latest.json, 11 assets in total. In-app updates still verify latest.json and .sig; the in-app size and SHA256 checks for model and engine ZIPs are unchanged. Assets of published releases stay as they are.
• Existing-model panel polish: the local-GGUF panel in the guide and on Download models now shares one left edge and spacing rhythm; the chosen path sits in a field; each quantization row keeps the name on the left and size + a quiet status dot on the right (color only for problems); MTP and the equivalent command collapse by default; long help text is shortened to one line.
• Use an existing model (#75): the guide's model step and the Download models page now offer "Use an existing model". Choose the folder with your GGUF files (or one shard); Manager groups them by model and quantization, checks that every shard is present and readable with a valid GGUF header and size, and has the official installer use them in place with --gguf-dir, without downloading the weights again. An existing MTP draft layer can be chosen (upstream setup.py has no --mtp setup flag, so it maps to --data-dir), and the equivalent START-HERE.bat command is shown. Renamed files with a supported quantization get a warning instead of being refused; OrcaRouter Uncensored IQ3_XXS uses the ORCA.md compatibility workflow; quantizations the engine cannot run are explained. A detected Strata runtime that is not registered yet can now be used to prepare models directly instead of showing "Choose a Strata runtime first".

Strata Manager v0.2.5

中文

• 跳过向导后仍可使用本地安装包（#74）：Strata 页面新增「使用已下载的安装包」入口。未安装时可选择本机已下载的 Strata 源码 ZIP 和（Windows）引擎 ZIP 安装，校验与安装流程与向导完全相同；运行环境需要补齐时同样可用；已安装时可通过「使用本地安装包」用官方引擎 ZIP 更新或重装当前版本，安装前按官方 SHA256 校验。

English

• Local packages after skipping the guide (#74): the Strata page now offers "Use downloaded packages". Without a runtime you can install from an already-downloaded Strata source ZIP and (on Windows) engine ZIP with the same validation and installation as the guide; this also works when the runtime needs repair. With a runtime installed, "Install from local package" updates or reinstalls the engine from an official engine ZIP, verified against the official SHA256 first.

Strata Manager v0.2.4

中文

• 多显卡显示（#69）：总览按「2× 显卡型号」显示同型号多卡，并逐卡列出显存、负载和温度；运行环境卡片不再只显示首张卡加「+1」。PCIe 链路注明来自 GPU 0。
• 大模型加载不再误报失败（#70）：冷加载超过 /load 请求超时（5 分钟）但引擎仍在加载时，继续等待模型就绪（最长 30 分钟，进程退出才判失败），总览显示「模型加载中（已 N 秒）」；启动失败记录在模型加载中或已就绪后自动隐藏，且不再把原始 JSON 显示给用户。

English

• Multiple GPUs (#69): the overview names identical cards as "2× model" with per-card VRAM, load and temperature; the environment strip no longer shows only the first card plus "+1". The PCIe link is labelled as GPU 0's.
• Slow model loads are no longer reported as failures (#70): when a cold load outlasts the 5-minute /load request timeout while the engine is still loading, startup keeps waiting for readiness (up to 30 minutes; only a process exit fails it) and the overview shows "Loading the model (N s so far)". Stale startup failures are hidden while the model loads or once it is ready, and raw JSON is never shown.

Strata Manager v0.2.3

中文

• 本地 / ZIP 引擎安装报「系统找不到指定的路径」（#65）：修复 v0.2.2 的回归——引擎解包目录已缩短为 e，但读取 BUILD.json 仍使用旧的 engine 目录，导致 HIP 等 ZIP 安装与更新在解包后立即报 os error 3。解包、读取与替换现统一使用同一目录。
• 虚拟显示器被识别为显卡（#66）：Windows 未检测到 NVIDIA 显卡时，不再把 Meta Virtual Monitor、Parsec、远程桌面、Microsoft 基本显示适配器等软件/虚拟显示设备列为 GPU。

English

• Local / ZIP engine install fails with "path not found" (#65): fixes a v0.2.2 regression where the engine was extracted into the shortened e directory but BUILD.json was still read from the old engine directory, so HIP and other ZIP installs/updates failed with os error 3 right after extraction. Extraction, reading and replacement now share one directory.
• Virtual displays detected as GPUs (#66): on Windows without an NVIDIA GPU, software/virtual display adapters such as Meta Virtual Monitor, Parsec, Remote Desktop and Microsoft Basic Display are no longer listed as GPUs.

Strata Manager v0.2.2

中文

• 本地 Strata 源码包识别（#63）：从 GitHub Release 下载的 Strata-<版本>.zip / Strata-v<版本>.zip（顶层目录不再仅限 Strata-<commit>/）可直接作为本地安装包通过校验。
• HIP / 深层引擎 ZIP 解包（#62）：引擎解包改为直接写入并缩短工作目录，Windows 使用长路径前缀；避免 tempfile + MoveFileEx 在深层 rocm/... 路径上报「系统找不到指定的路径」。路径过长时给出中文提示。

English

• Local Strata source archives (#63): GitHub Release Strata-<version>.zip / Strata-v<version>.zip trees are accepted as local install archives (not only Strata-<commit>/ folder names).
• HIP / deep engine ZIP extract (#62): engine unpack writes files directly with a shorter work directory and Windows long-path prefixes, avoiding tempfile + MoveFileEx “path not found” on deep rocm/... trees; overlong paths get a clear Chinese hint.

Strata Manager v0.2.1

中文

• Python 依赖误报修复：扫描时优先使用 Strata 启动脚本（run-*.bat / run-*.sh）里的私有 Python，冷启动探测等待放宽到约 20 秒，超时也不再偷偷换到系统解释器，避免 bat 能跑而 Manager 误标「需要修复」。
• 内存需求文案：模型「运行约需 X GB」改为峰值参考，并标明显存卸载后可更低；有 NVIDIA 显卡时，内存单独不够只警告为「偏紧」，不阻挡「启用配置」。
• 顶栏 CTA：本机已扫到模型但尚未启用服务时，顶栏由「尚未配置服务」改为「已发现模型，点启用」，并可一键跳到「我的模型」。
• 显存容量：服务未运行时，若 nvidia-smi 已给出容量，不再长期显示「未实测」，改为展示已知显存容量。
• 一键修复运行环境：缺 regex / jinja2 时，向 Strata 启动脚本使用的同一私有 Python 执行 pip install（不另造解释器）。
• 一键复制诊断信息：打包引擎版本、安装目录、Python 路径、缺失模块、近期日志摘要与本机 OS/GPU/RAM 到剪贴板，便于发给支持。

English

• Python dependency false “needs repair”: discovery prefers the private Python named by Strata launchers (run-*.bat / run-*.sh), allows ~20s for cold Windows probes, and no longer silently switches interpreters on timeout — so a working bat install is not flagged for repair.
• Memory need copy: the “needs about X GB” line is now an official peak figure and notes GPU offload can lower host RAM; with an NVIDIA GPU, short system RAM alone warns as tight and does not block Enable configuration.
• Title-bar CTA: when models are discovered but no service profile is enabled, the pill says “Models found — tap to enable” and jumps to My models.
• VRAM capacity: when the service is stopped but nvidia-smi already reported capacity, the UI shows that capacity instead of lingering “Not measured”.
• One-click repair runtime: missing regex / jinja2 are installed into the same private Python the Strata launchers use — no second interpreter.
• One-click copy diagnostics: engine version, install dir, Python path, missing modules, a short recent-log snippet, and OS/GPU/RAM as plain text for support.

Strata Manager v0.2.0

中文

• 一键提速（参数调优第一期）：调优页默认「一键提速」，按通用场景测速并扫描可提速参数，显示前后对比；结果另存为新配置，可回退。场景验收收进高级；短请求诊断入口已去掉；修复快速校准把 PCIe 专家比例写死从而关掉自动的问题。
• 排行榜：修复切换排序后一直显示「暂时连不上排行榜」——请求参数白名单补上排序与搜索；真正断网才显示连不上，其它错误显示具体原因；重试时恢复默认筛选与排序。
• 移除模型配置页的「资源适配建议」；首次安装向导与下载页的精简资源预估保留。
• Qwen3.6-35B-A3B：模型目录与安装流程支持引擎 0.1.45 的三档量化（UD-Q4_K_XL / UD-Q6_K_XL / Q8_0）；引擎过低时下载页禁用并提示。
• OrcaRouter 去审查模型：下载模型页可导入官方兼容的 IQ3_XXS（自动下载分片或选本地），后台兼容打包后出现在「我的模型」，标明社区兼容、仅文本。
• Unsloth 更高量化：在原有档位外增加 UD-Q5_K_XL、UD-Q6_K_XL、Q8_0（内存需求更高）。
• 首次安装：支持选择已下载的本地 Strata / 引擎安装包并跳过下载。
• 复制为新配置：已登记配置的副本会自动登记，可在配置列表中选用。
• 发布验签：补上 darwin-aarch64 与 DMG 资产校验，使含 Mac 实验包的 Release 能通过门禁并公开。
• macOS 第一期（实验）：CI 在 macos-15 / aarch64-apple-darwin 打出未公证 DMG 与 updater 用 .app.tar.gz（临时签名）。文档与官网标明实验包、Gatekeeper 限制，以及 DMG alone 不含 Metal 引擎（引擎仍待后续 ./setup.sh）。不提供 Intel 包。关联 #53。

English

• One-click speed boost (Tuning phase 1): Tuning defaults to one-click speedup for a general workload, compares before/after, and saves a new configuration you can roll back. Scenario checks move under Advanced; the short-request diagnostics entry is removed; quick calibration no longer pins PCIe expert share and disables auto.
• Leaderboard: fixing sort switches that stuck on “leaderboard unavailable” — allowlist now includes sort and search; only real network failures show unavailable; other errors show a concrete reason; retry resets filters and sort.
• Removed the Resource fit advice panel from model configuration; the compact estimate on onboarding and download remains.
• Qwen3.6-35B-A3B: catalog and install flow support engine 0.1.45 quants (UD-Q4_K_XL / UD-Q6_K_XL / Q8_0); downloads disable with a hint when the engine is too old.
• OrcaRouter uncensored: Download models can import the official-compatible IQ3_XXS (fetch shards or pick local files), run compat packaging in the background, and list the result under My models as community-compatible, text-only.
• Higher Unsloth quants: adds UD-Q5_K_XL, UD-Q6_K_XL and Q8_0 (higher memory needs).
• First-time install: choose an already-downloaded local Strata / engine archive and skip the download.
• Copy as new configuration: copies of registered configurations are registered automatically and selectable in the list.
• Release verification: allow darwin-aarch64 and DMG assets so a release that includes the experimental Mac packages can pass the gate and publish.
• macOS phase 1 (experimental): CI on macos-15 / aarch64-apple-darwin produces an un-notarized DMG and updater .app.tar.gz (ad-hoc signing). Docs and the website mark the package experimental, note Gatekeeper limits, and state that the DMG alone does not include a Metal engine (engine setup remains a later ./setup.sh phase). No Intel package. See #53.

Strata Manager v0.1.13

中文

• 并发支持：运行配置页新增“并发”卡片，按显存、内存和显卡数量给出建议：16 GB 单卡推荐会话缓存，32 GB 推荐同时处理 4 个请求，多卡按层切分；一键应用时会同时设置所需的配套参数。会检查 parallel 与 --batch 冲突（可一键修复）、--batch-groups 无法整除（拒绝保存），超过 262K 上下文时提示并发与长上下文冲突。新增 --trim-stage-weights；STRATA_BATCH_ATTN_ROWS 等实验性环境开关只在高级区域显示，并标明“实验性”。
• 并发阶梯测试：场景测试页可按 1 → N 个同时请求逐级测试，实时显示每级的总吞吐和单请求速度，结束后给出推荐并发数。结果只保存在本机，不上传排行榜。
• 槽位监控：始终显示全部列，新增生效槽数、总速度、排队数和每个槽位的速度趋势；监控页和状态栏会合计多路生成的速度。
• 调优页并发对比（可选）：把候选并发设置另存为新的配置版本，依次切换并跑阶梯测试，结束后恢复原配置，用表格和图表对比结果。切换配置时不会中断测试；在快速校准时，如果并发大于 1 会给出提示。
• 1M 上下文：上下文预设新增 384K 和 1M。超过 262,144 时，运行配置页自动写入 --rope-scaling yarn 和对应倍数（1M 为 4），改回 262K 以内时自动移除；手动设置的值不会被改动，倍数不对时给出提示并提供“改为官方设置”。页面会常显“实验性”说明，以及 KV 缓存和 RoPE 表的占用；资源预估改为计入 YaRN 倍数和 RoPE 表，不再显示“未知”。
• 参数调优改版：可以实时看到步骤、轮次、正在测的参数、tok/s、进度和预计剩余时间，并可取消。新增“快速校准”（默认模式，对应官方 tools/calibrate.py）：依次测预填充分块、PCIe 专家比例和 MTP 草稿阈值，只临时覆盖请求参数，不改配置、不重启，只有提升超过 3% 才推荐。结果默认另存为新配置版本，覆盖当前配置需要确认，并保留 .bak；日志随报告保存。运行配置新增 --prefill-pipe。
• 场景测试页改版：先选场景、开始测试，再看实时进度和分场景结果。运行时显示当前请求、首 token 时间、已用时间与预计剩余，以及逐请求散点图。结果包含 5 项关键指标（含 P50/P95/P99）、分场景柱状图和表格，请求样本带输出片段。去掉全部折叠区，原有选项和导出功能保留。
• 排行榜筛选与排序：应用内与官网的排行榜可按显存、内存、上下文、KV、并发、显卡数和专家缓存筛选，并可按 8K 解码、128K 解码、预填充或机器数排序；官网的筛选结果写入链接，可以分享。
• Benchmark 页移除顶部的“当前服务能力”面板；服务能力核验仍在“API 与外部助手”页。
• 官网和 README 更新了产品介绍，新增开始使用引导、排行榜和旧版本配置清理的截图，并重截了日常体验截图。

English

• Concurrency: Runtime configuration adds a Concurrency card that recommends settings from VRAM, memory and GPU count: conversation cache on a single 16 GB card, 4 requests at once on 32 GB, and a layer split across multiple GPUs. Applying a recommendation also sets the companion parameters it needs. The card flags a parallel / --batch conflict (with a one-click fix), refuses to save a --batch-groups value that does not divide evenly, and notes the conflict between concurrency and contexts past 262K. Adds --trim-stage-weights; experimental environment switches such as STRATA_BATCH_ATTN_ROWS appear only under Advanced and are labeled Experimental.
• Concurrency ladder: the scenario benchmark can step from 1 to N concurrent requests, showing total throughput and per-request speed live for each step, then recommend a concurrency. Results stay on this machine and are not uploaded to the leaderboard.
• Slot monitor: always shows every column, and adds active slots, total speed, waiting requests and a per-slot speed trend. Monitoring and the status bar add up speed across generating streams.
• Concurrency comparison in Tuning (optional): saves candidate concurrency settings as new configuration versions, switches through them running the ladder, restores the original configuration, and compares results in a table and chart. Switching configurations does not interrupt the run, and quick calibration shows a note when concurrency is above 1.
• 1M context: context presets add 384K and 1M. Past 262,144, Runtime configuration writes --rope-scaling yarn and the matching factor (4 for 1M), and removes them when you go back to 262K or less. Manual values are left alone; a wrong factor shows a note with “Use official settings”. An Experimental note stays visible with the KV cache and RoPE table sizes, and resource estimates now include the YaRN factor and RoPE table instead of showing Unknown.
• Tuning redesign: live step, round, parameter under test, tok/s, progress and time remaining, with cancel. New quick calibration (the default, matching upstream tools/calibrate.py) measures prefill pipe, PCIe expert share and MTP draft threshold by overriding request parameters only. It changes no configuration and needs no restart, and recommends a value only when it is more than 3% faster. Results are saved as a new configuration version by default; overwriting the current configuration needs confirmation and keeps a .bak. The log is saved with the report. Runtime configuration adds --prefill-pipe.
• Scenario benchmark redesign: choosing scenarios, starting, live progress and per-scenario results now come first. A run shows the current request, time to first token, elapsed and remaining time, and a per-request scatter chart. Results show five key metrics (including P50/P95/P99), a per-scenario bar chart and table, and output excerpts in the request samples. No folded sections; every existing option and export is kept.
• Leaderboard filters and sorting: the in-app and website leaderboards filter by VRAM, memory, context, KV, concurrency, GPU count and expert cache, and sort by 8K decode, 128K decode, prefill or machine count. Website filters are kept in the link so you can share them.
• Removed the Current service capabilities panel from the top of Benchmark; capability checks remain on the API & assistants page.
• The website and README have an updated product introduction, new screenshots of the getting-started guide, leaderboard and old-version configuration cleanup, and recaptured daily-experience screenshots.

Strata Manager v0.1.12

中文

• Benchmark 排行榜：在应用内的排行榜页预览将要上传的内容并同意后，每次测试完成会匿名上传结果，可随时关闭、撤回并删除已上传数据。排行榜按“同模型 + 同量化”分组对比，组内展示硬件与启动参数，并为本机推荐参数和模型、标注可信度；公开排行榜位于 strata-manager.cosfoc.com/leaderboard。上传数据带每台机器的签名与运行证据，异常结果会被剔除；应用内的排行榜数据只经后台读取，界面本身不直接联网。
• 首次使用引导：新增“开始使用”页，把检查环境、安装 Strata、选择模型、下载、启动并加载、第一次聊天串成 6 步；进度会保存，已完成的步骤按真实状态自动识别并跳过。选模型时按本机显存和内存推荐“模型 + 量化”，并参考排行榜的公开速度数据（只读）。也可以直接连接本机已运行的服务。总览页会显示未完成引导的续接入口，完成后可在设置中重新打开。
• 运行配置页改版：去掉顶部的资源估算区块，页头只保留标题和常用操作，资源适配建议收成一行提示（可展开查看依据）；主编辑区为“启动参数”，文件名、复制和删除放在卡片标题行；有未保存改动时底部保存栏会提示。“返回我的模型”并入面包屑。
• 多版本 Strata 配置管理：多个 Strata 版本共用同一份 Strata-data 权重时，各版本的配置不再混在一起。每份配置标注“Strata 0.1.xx”并按版本从新到旧排列；默认只显示当前版本，其他版本收进“其他版本 · N”，可用“显示全部版本”开关查看全部；内容相同的配置合并为一行并标注“另有 N 个版本相同”（只影响显示）。新增“清理旧版本配置”：按版本列出旧版本目录中的配置，勾选后移入回收区并记录原路径，可以恢复；正在使用的、当前版本的配置以及模型切片、mtp 等共享文件永远不会列出。
• 接入 API 弹窗精简：说明文字更短，验证细节（token 上限、限时、不执行 MCP、图片测试用图、不保存结果）收进帮助提示；验证结果改为逐步清单，带状态图标和耗时，失败原因显示在对应步骤下，后续步骤标“已跳过”，整体结果为一行状态，能力信息显示为标签。
• API 使用统计页改版：顶部一行显示请求、成功率、失败、输入 Token、输出 Token、生成速度六项关键指标；原来的六张小图合成一张主趋势图，用标签切换；新增请求结果环形图，调用来源表增加成功率与占比。原本折叠的调用来源、实时资源、请求明细与导出、统计说明、高级诊断改为直接显示，统计说明改为标签页；时间范围改为分段切换。
• 官网迁到 Cloudflare：官网与 Benchmark 排行榜现位于 strata-manager.cosfoc.com，main 更新后自动部署；旧的 github.io 地址与更新日志等旧链接会跳转到新域名。
• 修复：Windows 上打开运行配置页时不再弹出“找不到 cublas64_13.dll”系统错误框。读取引擎参数列表时会套用与启动服务相同的运行库路径和环境变量；仍缺少运行库时只在页面内显示一行提示，参数照常可编辑；后台进程遇到缺失 DLL 时直接以错误码退出，不再弹出需要点击的系统对话框。
• 修复：总览页显存条与内存条的色块和图例颜色保持一致；高占用改由大号数字颜色和 85% 阈值线提示。
• 移除运行环境页的“用官方脚本更新 Strata”；更新 Strata 继续使用版本检查与更新包（官方 GitHub、镜像或本地 ZIP）。

English

• Benchmark leaderboard: after previewing what will be uploaded on the in-app Leaderboard page and agreeing, each finished test uploads its result anonymously; you can turn this off, withdraw, and delete uploaded data at any time. Results are compared within the same model and quantization, showing hardware and launch parameters per group, with parameter and model recommendations for this machine and a confidence label. The public leaderboard is at strata-manager.cosfoc.com/leaderboard. Uploads carry a per-machine signature and run evidence, and outliers are removed; the app reads leaderboard data only through its background process and the interface never connects to the network directly.
• Getting started guide: a new Getting started page links checking the environment, installing Strata, choosing a model, downloading, starting and loading, and the first chat into six steps. Progress is saved, and steps that are already done are detected from real state and skipped. Model choice recommends a model and quantization for this machine's VRAM and memory, using public leaderboard speed data (read-only). You can also connect to a service already running on this machine. Overview shows a resume entry while the guide is unfinished, and Settings can reopen it later.
• Runtime configuration redesign: the resource estimate block at the top is gone; the header keeps the title and common actions, and resource advice becomes a one-line hint that expands to show its basis. The main editor is “Launch parameters”, with file name, copy and delete on the card title row, and the bottom save bar flags unsaved changes. “Back to My models” moves into the breadcrumb.
• Multi-version Strata configurations: when several Strata versions share one Strata-data weight folder, their configurations no longer pile up together. Each configuration is labeled “Strata 0.1.xx” and sorted newest first; only the current version is shown by default, other versions fold into “Other versions · N”, and a “Show all versions” toggle shows everything. Identical configurations collapse into one row marked “Same in N other versions” (display only). New “Clean up old version configurations” lists configurations in older version folders by version; selected files move to a recovery folder with their original paths recorded so they can be restored. Configurations in use, the current version's configurations, and shared files such as model shards and mtp are never listed.
• Simpler Connect API dialog: shorter copy, with verification details (token cap, time limit, no MCP execution, test image, nothing saved) moved into a help tooltip; results are now a step checklist with status icons and timings, failure reasons under the step, later steps marked Skipped, a one-line overall result and capability chips.
• API usage redesign: a top row shows six key metrics: requests, success rate, failures, input tokens, output tokens and generation speed. The six small charts become one main trend chart with tabs; a request outcome ring chart is added, and the caller table adds success rate and share. Callers, live resources, request details and export, statistics notes and advanced diagnostics are shown directly instead of folded, with notes as tabs; the time range is a segmented switch.
• Website on Cloudflare: the website and the Benchmark leaderboard now live at strata-manager.cosfoc.com and deploy automatically from main; old github.io addresses and changelog links redirect to the new domain.
• Fix: opening Runtime configuration on Windows no longer shows a “cublas64_13.dll was not found” system error. Reading the engine's parameter list now uses the same library paths and environment as starting the service; if a library is still missing, the page shows a one-line note and parameters stay editable, and background processes exit with an error code instead of a system dialog that waits for a click.
• Fix: VRAM and memory bar segments on Overview now match their legend colors; high usage is shown by the large figure's color and the 85% threshold line.
• Removed “Update Strata with the official script” from the runtime page; updating Strata continues through the version check and release packages (official GitHub, mirrors or a local ZIP).