v0.5.3

中文

维护版本，无应用行为变化：重新生成仓库内提交的 Wails TypeScript 绑定，使其与 v0.5.2 的 ServerConfig.remoteStart 字段及 GetLanAddresses 的最新文档注释保持一致（社区贡献 by @Vergillll）。发布产物在构建时本就会重新生成绑定，已安装用户不受影响；本版修正的是直接消费仓库内生成物的开发场景（类型检查 / IDE / 二次开发）。

English

Maintenance release with no behavior change: the committed Wails TypeScript bindings are regenerated to match v0.5.2's ServerConfig.remoteStart field and the refreshed GetLanAddresses doc comment (community contribution by @Vergillll). Release artifacts regenerate bindings at build time, so installed users were never affected; this fixes developer-facing consumption of the committed bindings (type-checking / IDE / downstream builds).

v0.5.2

中文

v0.5.2 带来「局域网远程对话」：手机与电脑端组成搭档，互相借力。

• 局域网远程对话（本版亮点） — 电脑端一键出示配对二维码，手机扫码即可导入连接：聊天页新增「对话目标」切换，模型列表、对话与模型卸载全部直达电脑端 llama-server，手机不再受限于本机算力。连接失败时给出明确指引（地址 / 服务 / 防火墙），并支持一键远程启动电脑端服务（需在电脑端开启「允许远程启动」并设置 API Key）。配对地址按可达性自动排序并剔除虚拟网卡，二维码默认出示真实局域网地址。
• 安卓远程连接修复（重要） — 修复安卓 WebView 将发往局域网地址的请求误判为「混合内容」而静默拦截的问题：此前手机端无法连接任何局域网地址（本机回环不受影响，故仅远程档受累）。同时安卓端明确为「仅本地推理、不对外服务」，设置界面相应收敛。
• 安全加固 — GGUF 元数据解析对嵌套数组递归加上深度与数量上限；控制面拒绝非本机绑定与跨站请求；远程启动仅在显式开启后暴露于局域网，且强制携带密钥（API Key 或控制令牌，常量时比较），不支持远程停止。
• 稳定性修复 — 下载中断重试改为从真实 .part 进度续传并自动恢复卡死任务；修复 GetLlamaCpp 数据竞争并使恢复的暂停任务可续传；任务胶囊支持队列中任务重试与完整链接手势；清理九项审计发现的低危问题。
• 工程质量 — CI 新增 Go 竞态检测 job；引入基于 mock 运行时的 Playwright 冒烟套件；聊天页完成组件化拆分（消息列表 / 输入区 / 参数面板 / 通知栈）。

English

v0.5.2 ships "LAN remote chat": your phone and the PC become a team.

• LAN remote chat (highlight) — the PC shows a pairing QR code and the phone scans it to connect: a chat-page target switcher lists the PC's models, and streaming chats plus model unloads go straight to the PC's llama-server, so the phone is no longer bounded by its own hardware. Connection failures explain themselves (address / service / firewall) and offer a one-tap remote start of the PC's service (enable "allow remote start" on the PC and set an API key first). Pairing addresses are ranked by reachability with virtual adapters excluded, so the QR shows the real LAN address.
• Android remote-connection fix (important) — the Android WebView misjudged requests to LAN addresses as mixed content and silently blocked them: the phone could never reach any LAN address (loopback was unaffected, so only the remote tier suffered). Android is also now explicitly local-inference-only and the settings UI reflects it.
• Security hardening — GGUF metadata parsing bounds nested-array recursion depth and count; the control plane rejects off-host and cross-site requests; remote start is exposed to the LAN only when explicitly enabled and always token-gated (API key or control token, constant-time compared); remote stop does not exist by design.
• Stability fixes — mid-stream download retries resume from the real .part size and stalled tasks recover automatically; a GetLlamaCpp data race is closed and restored paused tasks become resumable; the task dock supports queued-task retry and full link gestures; nine low-severity audit findings cleared.
• Engineering quality — CI gains a Go race-detector job; a Playwright smoke suite over the mock runtime is wired in; the Chat page is decomposed into focused components (message list / composer / params panel / notice stack).

v0.5.1

中文

v0.5.1 是一个修复版本，解决 v0.5.0 的两个问题：

• 安卓应用内更新修复（重要） — v0.5.0 的安卓端在应用内点击「立即安装」后，系统安装确认弹窗不会出现，应用闪一下回到首页、更新静默失败。本次重建了安装状态回传链路：修复状态 PendingIntent 的可变标志、补齐确认弹窗的转发逻辑、主界面改为单实例并处理状态重投递——弹窗现已正常弹出，界面不再被重置，安装失败信息也会显示在更新弹窗中。v0.5.0 安卓用户请在本发布页手动下载 APK 覆盖安装一次，此后应用内更新恢复正常（安装完成后需手动重新打开应用，属系统行为）。
• 设置页 API Key 状态修复 — API Key 为空时桌面端错误显示「已设置」；现在密钥为空（含纯空白）时，桌面与手机端统一显示「未设置（无鉴权）」，保存时同步按后端规则去除首尾空白。

English

v0.5.1 is a fix release addressing two v0.5.0 issues:

• Android in-app update fixed (important) — on v0.5.0, tapping "Install now" never raised the system confirmation dialog: the app flashed back to the home page and the update silently failed. The install-status handback has been rebuilt (mutable status PendingIntent, confirmation-intent forwarding, single-task MainActivity with status redelivery); the dialog now appears, the UI no longer resets, and install failures surface in the update modal. Android users on v0.5.0: sideload this release's APK manually once — in-app updates work again from this version on (after a completed install the app must be reopened manually; that is system behavior).
• Settings API-key status fixed — with an empty key the desktop label incorrectly showed "already set"; an empty (or whitespace-only) key now shows "not set (no authentication)" on both desktop and phone, with the input trimmed on save to match the backend's persisted value.

v0.5.0

中文

v0.5.0:品牌焕新为 MyLlama，安卓端正式发布，桌面端全新视觉，推理性能与安全全面加固。核心变化:

🌟 本版亮点

• 安卓端正式发布 — 全新手机端界面(底部导航、安全区与软键盘适配、灵动任务胶囊),直连模式单模型常驻、即开即用;一键调优自动按 big.LITTLE 性能核规划线程;支持应用内自更新(侧载分发,arm64)。下载、聊天与模型管理体验与桌面端一致。
• 桌面端全新视觉 — 玻璃拟态 + 紫渐变浮岛设计语言,桌面 / 平板 / 手机三档响应式,浅色 / 深色双主题;首页改为「问候 + 模型状态卡 + 快速上手清单」布局,AI 是否可用一目了然。
• 推理性能细节拉满 — 模型参数页新增 n-gram 投机解码(零额外显存,对代码 / 模板类文本显著加速);支持 q8_0 KV 缓存量化撑起更长上下文;Windows 全量 offload 自动钉定最优线程数并关闭上下文检查点,多花在 CPU 上的每个周期都省下来。

🔒 安全与稳定

• API Key 全链路 — 密钥经环境变量传递,进程列表与启动日志不再可见;聊天、模型管理与模型卸载全部携带鉴权;服务运行中修改密钥会明确提示「重启后生效」并可一键重启。
• 重启竞态修复 — API 页「重启」真正完成 停止 → 等待退出 → 启动,不再只停不起。
• 失效模型自愈 — 聊天页选中的模型被删除 / 改名时自动切换可用模型并轻提示,不再裸报 model not found。

🪟 Windows 体验

• 系统托盘 — 关闭窗口最小化到托盘,随时唤回主窗口或退出;无头模式下同样提供托盘图标。
• 安装器品牌迁移 — 全新 MyLlama 安装包自动检测旧版 Llama Desktop:配置、模型库与 llama.cpp 运行时整体迁入新目录并改写记录路径,旧版静默卸载。

📦 其他

• Linux 端 GPU 探测扩展:AMD / Intel 等非 NVIDIA 显卡可见并参与一键调优。
• 聊天新增生成预设(精准 / 均衡 / 创意 / 随机 + 自定义另存)。
• 应用内教程支持远程更新(不升级也能获取新章节),手机端新增阅读页。
• 更名 llama-cpp-desktop → MyLlama:仓库、安装器与状态文件全部启用新名,旧配置自动迁移;CI 新增安卓构建、模拟器冒烟与真实服务链 e2e。

English

v0.5.0: the rebrand to MyLlama, the Android release, a brand-new desktop look, and across-the-board inference and security hardening. Highlights:

🌟 Headlines

• Android is here — a brand-new phone UI (bottom navigation, safe-area & soft-keyboard handling, floating task capsule), direct-mode serving with a single resident model that's ready out of the box, auto-tune capped to big.LITTLE performance cores, and in-app self-updates (sideload, arm64). Downloads, chat and model management match the desktop experience.
• A brand-new desktop look — glassmorphism and violet-gradient floating islands across three responsive tiers (desktop / tablet / phone) with light & dark themes; the Home page now leads with a greeting, a model status card and a quick-start checklist.
• Inference performance, tuned to the last detail — n-gram speculative decoding (zero extra VRAM, a real speed-up for code / template-style text), q8_0 KV-cache quantization for longer contexts, and Windows full-offload plans that pin the optimal thread count and disable context checkpoints.

🔒 Security & stability

• API key end to end — the key is delivered via an environment variable, invisible in process lists and startup logs; chat, model listing and unload all carry bearer auth; saving a key while the service runs offers an explicit one-click restart.
• Restart race fixed — the API page restart truly completes stop → wait → start.
• Stale model self-heal — a deleted or renamed selected model auto-switches with a notice instead of a raw "model not found".

🪟 Windows experience

• System tray — closing the window minimizes to the tray; headless mode gets a tray icon too.
• Installer migration — the MyLlama installer detects a legacy Llama Desktop install, migrates config, model library and the llama.cpp runtime into the new directory (rewriting recorded paths) and silently uninstalls the old product.

📦 Also

• Linux GPU probing extended: AMD / Intel and other non-NVIDIA cards are now visible to one-click tuning.
• Chat gains generation presets (precise / balanced / creative / random + custom).
• The in-app tutorial updates remotely (new sections without upgrading) and gains a phone reader page.
• Renamed llama-cpp-desktop → MyLlama across repo, installer and state files (legacy configs migrate automatically); CI adds Android build + emulator smoke and real-service e2e tests.

v0.3.9

中文

v0.3.9:生成速度折线图回归与跨平台地基。核心变化:

• 生成速度折线图回归 — API 路由页 Token 速度卡恢复「近 60 秒生成速度」曲线,每秒采样、滚动保留 60 个点;模型管理页标题下补回副标题行,与其他页面版式一致。
• 应用路径按操作系统规范化 — Windows 行为完全不变;macOS/Linux 下配置、缓存与默认模型 / llama.cpp 目录归入用户配置目录(旧配置自动迁移),不再依赖进程工作目录,为 Linux/macOS 支持铺路。
• 响应式与平台地基 — 新增移动 / 平板 / 桌面三档断点体系与平台状态分类器(本版不改变任何现有界面),为适配手机、平板等不同尺寸做准备。

English

v0.3.9: the decode-speed chart returns, plus cross-platform groundwork. Highlights:

• Decode-speed chart restored — the Token speed card on the API Router page regains the "decode speed over last 60s" line chart (1s sampling, rolling 60 points); the Models page gets its subtitle line back, matching every other page.
• Per-OS app path resolution — Windows behavior is unchanged; on macOS/Linux the config, caches and default model / llama.cpp directories move under the per-user config directory (legacy config migrates automatically) instead of relying on the process working directory, paving the way for Linux/macOS support.
• Responsive and platform foundation — a three-tier breakpoint system (mobile / tablet / desktop) and a platform-state classifier land without changing any existing UI, preparing for phone- and tablet-sized screens.

v0.3.8

中文

v0.3.8:推理显卡选择、一键调优实测校准与实时预览、服务日志工程、无头控制面与内置帮助。核心变化:

• 推理显卡选择 — 设置页可指定服务运行在哪张显卡(按 nvidia-smi UUID 经 CUDA_VISIBLE_DEVICES 钉定);多卡机器上一键调优按所选显卡规划参数,规划卡与运行卡不再脱节。
• 调优实测校准 — 一键调优先实测本机内存带宽(按硬件指纹缓存、跨机自动作废),在「全量 offload 上下文局促」与「cpu-moe 大上下文」之间自动优选;模型设置页新增「深度实测」,用自带 llama-bench 实测当前方案的真实解码速度。
• 一键调优移入参数设置页 — 调优结果实时填充表单,不再是列表页一闪而过的摘要。
• CUDA 兼容性三态显示 — 首页 CUDA 兼容性区分「兼容 / 需 CUDA ≥12.8 / 已满足」,Blackwell 显卡装好达标运行时后不再误报警告。
• 服务日志工程 — llama-server 日志落盘文件并跟踪:GUI↔无头交接后日志不再丢失;ANSI 剥离与进度行节流(修复 Windows \r\r\n 幽灵空行);容量 200→2000 行;界面按游标增量拉取。
• 无头模式控制面 — API 路由模式下新增 127.0.0.1:1900 控制面(健康 / 状态 / 日志 / 停止):破坏性端点仅回环可达,可选 token 鉴权,绑定失败自动降级不影响启动。
• 优雅停止 — 停止服务先中断、宽限 5 秒后才强杀(Windows 行为不变);被交接的服务仍按原方式停止。
• 交接收养加固 — GUI↔无头切换认领服务前校验进程创建时间,防止 PID 复用导致的误认领与误杀。
• 崩溃安全写入 — 配置与交接文件原子写入(临时文件 + fsync + 原子换名),断电不再产生撕裂文件;版本三处一致性校验加入质量门。
• 内置帮助文档 — 侧栏新增「帮助」页:9 章节中英双语教程(快速上手 / 运行时 / 下载 / 模型 / API / 聊天 / 设置 / 无头模式 / 常见问题)。

English

v0.3.8: serving-GPU selection, measured auto-tune calibration with live preview, service-log engineering, a headless control plane and an in-app help center. Highlights:

• Serving-GPU selection — pick the GPU the service runs on in Settings (pinned via CUDA_VISIBLE_DEVICES by stable nvidia-smi UUID); the auto-tuner plans against the selected card on multi-GPU hosts.
• Measured tune calibration — one-click tuning measures the machine's RAM bandwidth first (cached per hardware fingerprint, rejected across machines) and picks between a cramped full-offload and a large-context cpu-moe plan; Model Settings gains "Deep benchmark" measuring the real decode speed of the saved plan via the bundled llama-bench.
• Auto-tune lives in Model Settings — tuning fills the settings form in real time instead of flashing a summary on the list.
• Three-state CUDA compatibility — the Home CUDA row distinguishes Compatible / Requires CUDA ≥12.8 / Satisfied, so Blackwell cards with a qualifying runtime stop raising false alarms.
• Service-log engineering — llama-server logs go to a file and are tailed: logs survive GUI↔headless handover; ANSI is stripped and progress redraws throttled (fixing Windows \r\r\n phantom rows); the ring grew to 2000 lines with cursor-based incremental fetch.
• Headless control plane — API-route mode gains a 127.0.0.1:1900 control plane (health / status / logs / stop): destructive endpoints are loopback-enforced with optional token auth; bind failure degrades startup-free.
• Graceful stop — stopping the service interrupts first and escalates to kill after a 5 s grace (Windows unchanged); handed-over servers keep their existing stop path.
• Handover adoption hardening — adoption verifies the process creation time, defending against PID reuse (no more adopting or killing an unrelated process).
• Crash-safe writes — config and handover files are written atomically (temp + fsync + atomic rename); a version-triple sync check joined the quality gate.
• In-app documentation — the sidebar gains a Help page: nine bilingual tutorial sections (quickstart / runtime / downloads / models / api / chat / settings / headless / faq).

v0.3.7

中文

v0.3.7:快速上手引导、官方 Hugging Face 下载源、可选 API Key 与无障碍修复。核心变化:

• 快速上手清单 — 首页新增三步引导卡(安装运行时 → 下载模型 → 启动服务),全部完成或手动关闭后不再显示。
• 官方 Hugging Face 下载源 — 模型下载源在 HF Mirror、ModelScope 之外新增「Hugging Face 官方」选项。
• ModelScope 修复 — 修复 ModelScope 搜索与下载不可用的问题。
• 可选 API Key — 设置页「服务访问范围」新增可选 API Key:非空时 llama-server 以 --api-key 启动,调用推理接口需携带 Bearer Token(无凭据请求返回 401),留空保持无认证;局域网暴露时可防止同网段无凭据调用。
• 下拉框键盘导航 — 全部下拉(模型设置页推理参数 + 聊天模型选择)支持键盘操作:方向键移动高亮、Enter 选择、Esc 关闭、Home/End 跳转,修复键盘与读屏用户无法操作下拉的问题。
• 模型说明渲染 — 下载详情页「模型说明」不再原样显示 README 中的 HTML/Markdown 标记,改为剥离内嵌 HTML 后渲染。
• 系统信息修正 — 修正 GPU Compute Capability 识别与系统信息缓存探测的问题。
• 启动与界面细节 — 补应用图标消除启动时 favicon.ico 404;首页卡片排版统一;API 页监控与 Token 布局修正。

English

v0.3.7: quick-start onboarding, official Hugging Face download source, optional API key and accessibility fixes. Highlights:

• Quick-start checklist — the Home page gains a three-step onboarding card (install the runtime → download a model → start the service) that hides once every step is done or the user dismisses it.
• Official Hugging Face source — model downloads gain an "official Hugging Face" option alongside HF Mirror and ModelScope.
• ModelScope fixed — ModelScope search and downloads are usable again.
• Optional API key — the Settings access-scope section gains an optional API key: when set, llama-server starts with --api-key and inference requests must carry the bearer token (unauthenticated requests get 401); leaving it empty keeps today's no-authentication behavior — useful to stop credential-free calls when the service is exposed to the LAN.
• Dropdown keyboard navigation — all dropdowns (ModelSettings inference params + the chat model picker) are fully keyboard-operable: arrow keys move the highlight, Enter selects, Escape closes, Home/End jump — fixing keyboard and screen-reader accessibility.
• Model description rendering — the download detail page no longer shows raw HTML/Markdown source in the description; embedded HTML is stripped and the remainder is rendered.
• System info fixes — corrected GPU compute capability detection and cached system-info probing.
• Startup & UI polish — added the app icon to remove the per-startup favicon.ico 404; Home card typography unified; API page monitor and token layout fixed.

v0.3.6

Full Changelog: v0.3.5...v0.3.6

v0.3.5

中文

v0.3.5:API 路由页布局重构、llama.cpp 下载修复与下载管理分页。核心变化:

• llama.cpp 下载修复 — 适配 llama.cpp 新发布策略(二进制只随每夜 prerelease 发布,latest 仅剩标记文件):自动回退到 release 列表,选取最新带平台构建的每夜版本下载。
• 下载取消修复 — 元数据获取与传输请求全部绑定取消上下文,点击「取消」立即中断在途请求;修复「暂停后取消」状态卡死在 paused、取消与继续按钮双双失灵的问题。
• 下载管理分页 — 下载任务与下载历史拆成两个页签,完成的历史不再混入任务列表;弹窗固定高度、切换页签尺寸不变,长列表内部滚动、页签吸顶。
• 按钮用词统一 — llama.cpp 下载的「停止」改为「取消」,与模型下载、应用更新一致。
• API 路由页布局 — 宽度上限提至 1920px,全屏几乎无左右留白;大窗口下一个参数一行、间距随窗口增大,监控区离底悬浮不再贴边;矮窗口自动紧凑模式(处理器/内存并排、隐藏次要说明),最小窗口回到 900×600 仍完整无滚动。

English

v0.3.5: API Router layout rework, llama.cpp download fixes and a tabbed download manager. Highlights:

• llama.cpp download fixed — adapts to llama.cpp's new release strategy (binaries ship only in nightly prereleases while /releases/latest is an asset-less marker): the app falls back to the release list and downloads the newest nightly carrying a platform build.
• Download cancel fixed — release-metadata fetches and transfer requests are all bound to the cancel context, so Cancel aborts in-flight requests immediately; also fixed stop-from-paused stranding the state machine in paused with Cancel and Resume both dead.
• Tabbed download manager — download tasks and download history are separate tabs, so finished downloads no longer mix into the task list; the modal keeps a fixed height across tab switches, long lists scroll internally with a sticky tab bar.
• Consistent wording — the llama.cpp download stop button now reads Cancel, matching model downloads and app updates.
• API Router layout — page width cap raised to 1920px (nearly no side margins when maximized); tall windows show one metric per row with spacing that grows with the window and the monitor area floating above the bottom edge; short windows switch to a compact mode (CPU + memory side by side, secondary captions hidden) so the 900x600 minimum stays fully visible without scrolling.

v0.3.4

中文

v0.3.4:API 路由页监控区精简与固定布局。核心变化:

• 移除 Token 速度折线图 — API 路由页 Token 卡只保留提示词处理速度与生成速度两个数值指标,近 60 秒生成速度历史曲线及图表辅助代码一并删除。
• 监控列固定无滚动 — 系统监控与 Token 速度两张卡始终完整填满右列,不再出现内部滚动条;启动 / 停止服务时卡片高度保持不变,未启动时仅显示占位提示。
• 页面底部收紧 — 监控网格与窗口底部的间距从约 80px 收窄到 24px,日志控制台与监控区可视高度增加。
• 下载搜索示例更新 — 模型下载页搜索框占位示例改为 LLM 模型(如 Qwen3.8-27B、Qwen3.6-35B-A3B)。
• 窗口最小尺寸提高 — 主窗口最小尺寸从 900×600 提高到 900×800,与默认高度一致,矮窗口不再裁剪监控区内容。
• 安装路径记忆修复 — 归属信息变更使安装器卸载注册表键名改变,覆盖安装丢失自定义路径回落默认目录;现在按「当前键 → v0.3.x 旧键 → llama-gui 旧键」顺序读回上次安装路径,并清理被取代的旧键避免重复条目。

English

v0.3.4: streamlined, fixed-layout monitor on the API Router page. Highlights:

• Token speed chart removed — the Token card keeps only the two numeric metrics (prompt processing / generation speed); the 60-second decode-speed history chart and its helpers are gone.
• Scroll-free monitor column — the System monitor and Token speed cards always fill the right column completely with no internal scrollbar; card heights stay identical whether the service is running or stopped, with only the placeholder shown before start.
• Tighter bottom spacing — the monitor grid ends 24px above the window bottom (was ~80px), giving the log console and monitor more visible height.
• Download search examples — the model download page's search placeholder now shows LLM examples (e.g. Qwen3.8-27B, Qwen3.6-35B-A3B).
• Larger minimum window — the main window's minimum size rises from 900×600 to 900×800 (matching the default height), so short windows no longer clip the monitor cards.
• Install path memory fixed — an attribution-driven rename of the uninstall registry key made updates lose the remembered custom path and fall back to the default directory; the installer now reads the previous InstallLocation from the current key, then the v0.3.x-era key, then the pre-rename llama-gui key, and removes the superseded entry.