// Hearth · i18n (English fallback + zh-CN + zh-TW) // 设计：t(英文原串) 查表；缺失键回退英文原串（零破坏）。 // 语言：en 🇬🇧 / zh-CN 🇨🇳 / zh-TW 🇹🇼。localStorage 持久化 aim.lang。 (() => { const LANGS = [ { code: "en", flag: "🇬🇧", label: "English" }, { code: "zh-CN", flag: "🇨🇳", label: "简体中文" }, { code: "zh-TW", flag: "🇹🇼", label: "繁體中文" }, ]; // 字典：key = 英文原串。en 用原串本身（表里不重复列）。 const DICT = { "zh-CN": { // nav "Overview": "总览", "Cluster": "集群", "Nodes": "节点", "Models": "模型", "Telemetry": "遥测", "Fabric": "网络", "Search": "搜索", "nodes online": "节点在线", "OFFLINE": "离线", "reconnecting…": "重连中…", "mock data · no backend": "模拟数据 · 无后端", "LIVE": "实时", "MOCK": "模拟", // hero "Home": "家庭", "Compute Center": "智算中心", "A unified telemetry surface for your home AI compute cluster — auto-discovers models, surfaces real metrics from vLLM / llama.cpp / LiteLLM, honestly labels what backends don't expose. Every token, every watt — in real time.": "自托管的家庭 AI 算力可观测性面板 —— 自动发现模型, 从 vLLM / llama.cpp / LiteLLM 取真实指标, 后端不暴露的诚实标空。每一个 token, 每一瓦特, 实时。", "UPTIME": "运行时长", "REQUESTS": "请求数", "TOKENS SERVED": "已服务 Token", "SCRAPE": "抓取", // section eyebrows "Cluster · live telemetry": "集群 · 实时遥测", "Models · LiteLLM gateway": "模型 · LiteLLM 网关", "Telemetry · alerts": "遥测 · 告警", "Fabric · network": "网络 · 拓扑", // section h2 (拆分普通 + 两段) "Real-time pulse of the ": "实时脉搏 · ", "entire fabric.": "整个集群", "{n} machines. ": "{n} 台机器 · ", "Each a citizen.": "各为一员", "One endpoint. ": "一个入口 · ", "Every model.": "所有模型", "Signal, not ": "只见信号 · ", "noise.": "不见噪声", "The wires ": "万物 · ", "between everything.": "互联之线", // ledes "Aggregated utilization, throughput, and thermals across all {n} nodes — sampled at 1.2 s, retained for 7 days in Prometheus, surfaced here as a single coherent picture.": "{n} 节点的利用率、吞吐与热数据聚合 —— 1.2 秒采样，Prometheus 留存 7 天，在此汇成一张连贯全景。", "Each host runs the LiteLLM gateway, an inference engine, or both. Click a node for the full forensic view.": "每台主机跑 LiteLLM 网关、推理引擎,或两者皆有。点击任一节点查看完整取证视图。", "The gateway host peers with every inference node — link intensity & pulse are driven by real node-exporter throughput. Inter-node fabric (RDMA/InfiniBand) is physical topology only (throughput not instrumented by default).": "网关主机与各推理节点互联 —— 链路强度/脉冲由真实 node-exporter 吞吐驱动;节点间高速 fabric(RDMA/InfiniBand)为物理拓扑,默认不插桩吞吐。", // cards / labels "Resource pools": "资源池", "Aggregate capacity vs. live usage": "总容量 vs 实时用量", "Hosted models": "承载模型", "Running services": "运行服务", "System": "系统", "Quick actions": "快捷操作", "LiteLLM Gateway": "LiteLLM 网关", "Model": "模型", "Throughput": "吞吐", "State": "状态", "Accelerator activity": "加速器活跃度", "VRAM": "显存", // footer "Prometheus 2.55 · DCGM 3.3 · LiteLLM 1.52 · scrape 15 s · retention 7 d": "Prometheus 2.55 · DCGM 3.3 · LiteLLM 1.83 · 抓取 15 秒 · 留存 7 天", "All inference is routed through LiteLLM on ": "全部推理统一经 LiteLLM 路由 · ", " — OpenAI-compatible API, smart routing, fallbacks, cost & token accounting. Cold models spin up on demand to fit the VRAM budget.": " —— OpenAI 兼容 API、智能路由、回退、成本与 Token 计量。冷模型按需拉起以适配显存预算。", "Every request that lands at the gateway, every anomaly the rules engine catches — surfaced as a quiet, structured stream. No paging unless something actually needs you.": "每一个到达网关的请求、规则引擎捕获的每一处异常 —— 以安静、结构化的流呈现。除非真的需要你，否则不打扰。", "Models · LiteLLM gateway": "模型 · LiteLLM 网关", // PulseCards "Token throughput": "Token 吞吐", "tokens / sec · all models": "tokens/秒 · 全部模型", "Cluster power draw": "集群功耗", "watts · all PSUs": "瓦特 · 全部电源", "KV-cache pressure": "KV 缓存压力", "% utilization, weighted mean": "% 利用率 · 加权均值", // node detail "forensic view": "取证视图", "Accelerator activity · 60 ticks": "加速器活跃度 · 近 60 拍", "GPU temp": "GPU 温度", "NIC temp": "网卡温度", "Power draw": "功耗", "Disk usage": "磁盘占用", "Net In": "入向流量", "Net Out": "出向流量", "Hosted models": "承载模型", "Running services": "运行服务", "Kernel": "内核", "Driver": "驱动", "Net": "网络", "Hardware temps": "硬件温度", "Hardware sensors": "硬件传感器", "Fans": "风扇", "+{n} more": "还有 {n} 项", "— no temperature data (node unmanaged) —": "— 无温度数据（节点未纳管）—", "GPU pending · maintenance window": "GPU 待维护窗口纳管", "Open SSH session": "打开 SSH 会话", "Drain & cordon": "排空并隔离", "Restart daemons": "重启守护进程", "Pin model…": "固定模型…", "Run nvidia-smi dmon": "运行 nvidia-smi dmon", "Reboot": "重启节点", // model detail / table "Throughput · last 32 ticks": "吞吐 · 近 32 拍", "TTFT (first token)": "TTFT（首 token）", "TPOT (per token)": "TPOT（每 token）", "Concurrency · running": "并发 · 执行中", "running": "执行中", "queued": "排队中", "Requests / sec": "请求/秒", "Route · ": "路由 · ", "TTFT / TPOT": "TTFT / TPOT", "KV-cache": "KV 缓存", "P50 / P95 / P99": "P50 / P95 / P99", "Placement · framework": "部署 · 框架", "no live metrics source": "无实时指标源", "vLLM not exposed": "vLLM 未暴露", "vLLM + llama.cpp + SGLang /metrics": "vLLM + llama.cpp + SGLang /metrics", "Live throughput": "实时吞吐", "llama.cpp /metrics does not expose TTFT / KV% / e2e histograms · shown as — honestly": "llama.cpp /metrics 未暴露 TTFT / KV% / e2e 直方图 · 诚实显—，不伪造", "TPOT / KV-cache / error-rate: this vLLM build does not expose them → honestly left blank, not faked.": "TPOT / KV缓存 / 错误率：此 vLLM 构建 /metrics 未暴露 → 诚实留空，不伪造。", "backend has no Prometheus /metrics endpoint. No live-metrics source → honestly marked, not faked. Model still serves normally via LiteLLM route": "后端无 Prometheus /metrics 端点。实时指标无来源 → 诚实标注，不伪造。模型经 LiteLLM 路由正常可用", // gateway stats "Catalog models": "目录模型", "with live metrics": "有实时指标", "Live metric sources": "实时指标源", "vLLM native /metrics": "vLLM 原生 /metrics", "vLLM throughput": "vLLM 吞吐", "t/s · measured sum": "t/s · 实测合计", "LiteLLM metrics": "LiteLLM 指标", "enterprise": "企业版", "OSS not exposed · documented": "OSS 不暴露 · 已记录", "healthy": "健康", "streaming": "流式传输中", // filters / states "All": "全部", "RTX": "独显", "DGX": "统一", "Active": "活跃", "Serving": "服务中", "Idle": "空闲", "Chat": "对话", "serving": "服务中", "idle": "空闲", "no-metrics": "无指标", "online": "在线", "stopped": "已停", "loading": "加载中", "cold": "冷", "down": "离线", // cmdk "Search nodes, models, sections…": "搜索节点、模型、板块…", "Section · top of page": "板块 · 页面顶部", "Section · aggregate telemetry": "板块 · 聚合遥测", "Section · {n} machines": "板块 · {n} 台机器", "Section · LiteLLM gateway": "板块 · LiteLLM 网关", "Section · alerts & log stream": "板块 · 告警与日志流", "Section · network topology": "板块 · 网络拓扑", "No results": "无结果", // tweaks "Tweaks": "偏好设置", "Appearance": "外观", "Theme": "主题", "Dark": "深色", "Light": "浅色", "Accent": "强调色", "Density": "密度", "Compact": "紧凑", "Regular": "标准", "Comfy": "宽松", "Live data": "实时数据", "Source": "数据源", "Auto": "自动", "Live": "实时", "Mock": "模拟", "Tick": "刷新间隔", "Motion / animations": "动效 / 动画", "Shortcut": "快捷键", "Press": "按", "anywhere to open the command palette.": "可在任意处打开命令面板。", "Click any node or model row for the forensic view.": "点击任意节点或模型行进入取证视图。", "Framework": "框架", "Port": "端口", "Quant": "量化", "Context": "上下文", "Placement": "部署", "Open Playground": "打开 Playground", "Restart": "重启", "Logs": "日志", "Close": "关闭", "— No models pinned to this node —": "— 无模型固定到此节点 —", "ONLINE": "在线", "OFFLINE": "离线", "Power": "功率", "maint. pending": "待维护纳管", "LiteLLM request log · OSS Postgres": "LiteLLM 请求日志 · OSS Postgres", "East-West · 200 GbE ConnectX-7 RDMA (live)": "东西向 · 200 GbE ConnectX-7 RDMA（真实）", "Per-node": "各节点", "heatmap": "热力图", "Accelerator activity · last 60 ticks · 0–100%": "加速器活跃度 · 近 60 拍 · 0–100%", "VRAM usage · last 60 ticks · 0–100%": "显存占用 · 近 60 拍 · 0–100%", "East-West · 200 GbE ConnectX-7 (topology only, throughput not instrumented)": "东西向 · 200 GbE ConnectX-7（仅拓扑，吞吐未插桩）", "North-South pulse ∝ real network throughput": "南北向脉冲 ∝ 真实网络吞吐", "Hearth": "Hearth · 家庭算力监控", "Home AI Compute Center": "家庭智算中心", "Home AI Compute Monitor": "家庭 AI 算力监控", // lang switcher "Language": "语言", // ── Training section ── "Training": "训练", "Training · distributed": "训练 · 分布式", "The long ": "漫长的", "run.": "训练。", "Substrate health for the 4×GB10 training cluster — per-rank GPU, unified-memory headroom, CX-7 RoCE collective bandwidth, and a silent-stall heuristic. Read-only from obs; zero training impact.": "4×GB10 训练集群的底座健康 —— 逐卡 GPU、统一内存余量、CX-7 RoCE 集合通信带宽,以及静默卡死启发式检测。只读自 obs,对训练零影响。", "Ranks active": "活跃 rank", "util ≥ 50%": "利用率 ≥ 50%", "avg util": "平均利用率", "RoCE Σ rx+tx": "RoCE Σ 收+发", "GPU power Σ": "GPU 功率 Σ", "healthy": "健康", "idle": "空闲", "GPU fault": "GPU 故障", "stall suspected": "疑似卡死", "skew": "偏斜", "util": "利用率", "straggler": "落后节点", "stall?": "卡死?", "mem": "内存", "comm stalled?": "通信卡死?", "train loss": "训练 loss", "step": "步", "mean accept": "平均接受", "grad norm": "梯度范数", "lr": "学习率", "logging": "记录中", "write stalled?": "写入停滞?", "source": "来源", "refresh ~12s": "刷新 ~12s", "MFU pending GB10 bf16 peak FLOPs.": "MFU 待 GB10 bf16 峰值 FLOPs。", "Loss · step / ETA · accept rate — awaiting a training signal source (auto-detects JSON / Prometheus / TensorBoard from the leader). Not fabricated. See docs/training-observability.md.": "Loss · 步 / ETA · 接受率 —— 等待训练信号源(从 leader 自动探测 JSON / Prometheus / TensorBoard)。不伪造。见 docs/training-observability.md。", }, "zh-TW": { "Overview": "總覽", "Cluster": "叢集", "Nodes": "節點", "Models": "模型", "Telemetry": "遙測", "Fabric": "網路", "Search": "搜尋", "nodes online": "節點在線", "OFFLINE": "離線", "reconnecting…": "重連中…", "mock data · no backend": "模擬資料 · 無後端", "LIVE": "即時", "MOCK": "模擬", "Home": "家庭", "Compute Center": "智算中心", "A unified telemetry surface for your home AI compute cluster — auto-discovers models, surfaces real metrics from vLLM / llama.cpp / LiteLLM, honestly labels what backends don't expose. Every token, every watt — in real time.": "自架的家庭 AI 算力可觀測性面板 —— 自動發現模型, 從 vLLM / llama.cpp / LiteLLM 取真實指標, 後端不暴露的誠實標空。每一個 token, 每一瓦特, 即時。", "UPTIME": "運行時長", "REQUESTS": "請求數", "TOKENS SERVED": "已服務 Token", "SCRAPE": "擷取", "Cluster · live telemetry": "叢集 · 即時遙測", "Models · LiteLLM gateway": "模型 · LiteLLM 閘道", "Telemetry · alerts": "遙測 · 告警", "Fabric · network": "網路 · 拓撲", "Real-time pulse of the ": "即時脈搏 · ", "entire fabric.": "整個叢集", "{n} machines. ": "{n} 台機器 · ", "Each a citizen.": "各為一員", "One endpoint. ": "一個入口 · ", "Every model.": "所有模型", "Signal, not ": "只見訊號 · ", "noise.": "不見雜訊", "The wires ": "萬物 · ", "between everything.": "互聯之線", "Aggregated utilization, throughput, and thermals across all {n} nodes — sampled at 1.2 s, retained for 7 days in Prometheus, surfaced here as a single coherent picture.": "{n} 節點的利用率、吞吐與熱數據聚合 —— 1.2 秒取樣，Prometheus 留存 7 天，在此匯成一張連貫全景。", "Each host runs the LiteLLM gateway, an inference engine, or both. Click a node for the full forensic view.": "每台主機跑 LiteLLM 閘道、推論引擎,或兩者皆有。點擊任一節點查看完整取證視圖。", "The gateway host peers with every inference node — link intensity & pulse are driven by real node-exporter throughput. Inter-node fabric (RDMA/InfiniBand) is physical topology only (throughput not instrumented by default).": "閘道主機與各推論節點互聯 —— 鏈路強度/脈衝由真實 node-exporter 吞吐驅動;節點間高速 fabric(RDMA/InfiniBand)為實體拓撲,預設不插樁吞吐。", "Resource pools": "資源池", "Aggregate capacity vs. live usage": "總容量 vs 即時用量", "Hosted models": "承載模型", "Running services": "執行服務", "System": "系統", "Quick actions": "快捷操作", "LiteLLM Gateway": "LiteLLM 閘道", "Model": "模型", "Throughput": "吞吐", "State": "狀態", "Accelerator activity": "加速器活躍度", "VRAM": "顯存", "Prometheus 2.55 · DCGM 3.3 · LiteLLM 1.52 · scrape 15 s · retention 7 d": "Prometheus 2.55 · DCGM 3.3 · LiteLLM 1.83 · 擷取 15 秒 · 留存 7 天", "All inference is routed through LiteLLM on ": "全部推理統一經 LiteLLM 路由 · ", " — OpenAI-compatible API, smart routing, fallbacks, cost & token accounting. Cold models spin up on demand to fit the VRAM budget.": " —— OpenAI 相容 API、智慧路由、回退、成本與 Token 計量。冷模型按需拉起以適配顯存預算。", "Every request that lands at the gateway, every anomaly the rules engine catches — surfaced as a quiet, structured stream. No paging unless something actually needs you.": "每一個到達閘道的請求、規則引擎捕獲的每一處異常 —— 以安靜、結構化的流呈現。除非真的需要你，否則不打擾。", "Token throughput": "Token 吞吐", "tokens / sec · all models": "tokens/秒 · 全部模型", "Cluster power draw": "叢集功耗", "watts · all PSUs": "瓦特 · 全部電源", "KV-cache pressure": "KV 快取壓力", "% utilization, weighted mean": "% 使用率 · 加權均值", "forensic view": "取證視圖", "Accelerator activity · 60 ticks": "加速器活躍度 · 近 60 拍", "GPU temp": "GPU 溫度", "NIC temp": "網卡溫度", "Power draw": "功耗", "Disk usage": "磁碟佔用", "Net In": "入向流量", "Net Out": "出向流量", "Hosted models": "承載模型", "Running services": "執行服務", "Kernel": "核心", "Driver": "驅動", "Net": "網路", "Hardware temps": "硬體溫度", "Hardware sensors": "硬體感測器", "Fans": "風扇", "+{n} more": "還有 {n} 項", "— no temperature data (node unmanaged) —": "— 無溫度資料（節點未納管）—", "GPU pending · maintenance window": "GPU 待維護窗口納管", "Open SSH session": "開啟 SSH 工作階段", "Drain & cordon": "排空並隔離", "Restart daemons": "重啟守護程序", "Pin model…": "固定模型…", "Run nvidia-smi dmon": "執行 nvidia-smi dmon", "Reboot": "重啟節點", "Throughput · last 32 ticks": "吞吐 · 近 32 拍", "TTFT (first token)": "TTFT（首 token）", "TPOT (per token)": "TPOT（每 token）", "Concurrency · running": "並發 · 執行中", "running": "執行中", "queued": "排隊中", "Requests / sec": "請求/秒", "Route · ": "路由 · ", "TTFT / TPOT": "TTFT / TPOT", "KV-cache": "KV 快取", "P50 / P95 / P99": "P50 / P95 / P99", "Placement · framework": "部署 · 框架", "no live metrics source": "無即時指標來源", "vLLM not exposed": "vLLM 未暴露", "vLLM + llama.cpp + SGLang /metrics": "vLLM + llama.cpp + SGLang /metrics", "Live throughput": "即時吞吐", "llama.cpp /metrics does not expose TTFT / KV% / e2e histograms · shown as — honestly": "llama.cpp /metrics 未暴露 TTFT / KV% / e2e 直方圖 · 誠實顯—,不偽造", "TPOT / KV-cache / error-rate: this vLLM build does not expose them → honestly left blank, not faked.": "TPOT / KV快取 / 錯誤率：此 vLLM 建置 /metrics 未暴露 → 誠實留空，不偽造。", "backend has no Prometheus /metrics endpoint. No live-metrics source → honestly marked, not faked. Model still serves normally via LiteLLM route": "後端無 Prometheus /metrics 端點。即時指標無來源 → 誠實標註，不偽造。模型經 LiteLLM 路由正常可用", "Catalog models": "目錄模型", "with live metrics": "有即時指標", "Live metric sources": "即時指標來源", "vLLM native /metrics": "vLLM 原生 /metrics", "vLLM throughput": "vLLM 吞吐", "t/s · measured sum": "t/s · 實測合計", "LiteLLM metrics": "LiteLLM 指標", "enterprise": "企業版", "OSS not exposed · documented": "OSS 不暴露 · 已記錄", "healthy": "健康", "streaming": "串流中", "All": "全部", "RTX": "独显", "DGX": "统一", "Active": "活躍", "Serving": "服務中", "Idle": "閒置", "Chat": "對話", "serving": "服務中", "idle": "閒置", "no-metrics": "無指標", "online": "在線", "stopped": "已停", "loading": "載入中", "cold": "冷", "down": "離線", "Search nodes, models, sections…": "搜尋節點、模型、區塊…", "Section · top of page": "區塊 · 頁面頂部", "Section · aggregate telemetry": "區塊 · 聚合遙測", "Section · {n} machines": "區塊 · {n} 台機器", "Section · LiteLLM gateway": "區塊 · LiteLLM 閘道", "Section · alerts & log stream": "區塊 · 告警與日誌流", "Section · network topology": "區塊 · 網路拓撲", "No results": "無結果", "Tweaks": "偏好設定", "Appearance": "外觀", "Theme": "主題", "Dark": "深色", "Light": "淺色", "Accent": "強調色", "Density": "密度", "Compact": "緊湊", "Regular": "標準", "Comfy": "寬鬆", "Live data": "即時資料", "Source": "資料來源", "Auto": "自動", "Live": "即時", "Mock": "模擬", "Tick": "刷新間隔", "Motion / animations": "動效 / 動畫", "Shortcut": "快捷鍵", "Press": "按", "anywhere to open the command palette.": "可在任意處開啟命令面板。", "Click any node or model row for the forensic view.": "點擊任一節點或模型列進入取證視圖。", "Framework": "框架", "Port": "連接埠", "Quant": "量化", "Context": "上下文", "Placement": "部署", "Open Playground": "開啟 Playground", "Restart": "重啟", "Logs": "日誌", "Close": "關閉", "— No models pinned to this node —": "— 無模型固定到此節點 —", "ONLINE": "在線", "OFFLINE": "離線", "Power": "功率", "maint. pending": "待維護納管", "LiteLLM request log · OSS Postgres": "LiteLLM 請求日誌 · OSS Postgres", "East-West · 200 GbE ConnectX-7 RDMA (live)": "東西向 · 200 GbE ConnectX-7 RDMA（真實）", "Per-node": "各節點", "heatmap": "熱力圖", "Accelerator activity · last 60 ticks · 0–100%": "加速器活躍度 · 近 60 拍 · 0–100%", "VRAM usage · last 60 ticks · 0–100%": "顯存佔用 · 近 60 拍 · 0–100%", "East-West · 200 GbE ConnectX-7 (topology only, throughput not instrumented)": "東西向 · 200 GbE ConnectX-7（僅拓撲，吞吐未插樁）", "North-South pulse ∝ real network throughput": "南北向脈衝 ∝ 真實網路吞吐", "Hearth": "Hearth · 家庭算力监控", "Home AI Compute Center": "家庭智算中心", "Home AI Compute Monitor": "家庭 AI 算力監控", "Language": "語言", // ── Training section ── "Training": "訓練", "Training · distributed": "訓練 · 分散式", "The long ": "漫長的", "run.": "訓練。", "Substrate health for the 4×GB10 training cluster — per-rank GPU, unified-memory headroom, CX-7 RoCE collective bandwidth, and a silent-stall heuristic. Read-only from obs; zero training impact.": "4×GB10 訓練叢集的底座健康 —— 逐卡 GPU、統一記憶體餘量、CX-7 RoCE 集合通訊頻寬,以及靜默卡死啟發式偵測。唯讀自 obs,對訓練零影響。", "Ranks active": "活躍 rank", "util ≥ 50%": "利用率 ≥ 50%", "avg util": "平均利用率", "RoCE Σ rx+tx": "RoCE Σ 收+發", "GPU power Σ": "GPU 功率 Σ", "healthy": "健康", "idle": "空閒", "GPU fault": "GPU 故障", "stall suspected": "疑似卡死", "skew": "偏斜", "util": "利用率", "straggler": "落後節點", "stall?": "卡死?", "mem": "記憶體", "comm stalled?": "通訊卡死?", "train loss": "訓練 loss", "step": "步", "mean accept": "平均接受", "grad norm": "梯度範數", "lr": "學習率", "logging": "記錄中", "write stalled?": "寫入停滯?", "source": "來源", "refresh ~12s": "刷新 ~12s", "MFU pending GB10 bf16 peak FLOPs.": "MFU 待 GB10 bf16 峰值 FLOPs。", "Loss · step / ETA · accept rate — awaiting a training signal source (auto-detects JSON / Prometheus / TensorBoard from the leader). Not fabricated. See docs/training-observability.md.": "Loss · 步 / ETA · 接受率 —— 等待訓練訊號源(從 leader 自動偵測 JSON / Prometheus / TensorBoard)。不偽造。見 docs/training-observability.md。", }, }; const saved = localStorage.getItem("aim.lang"); const navZh = (navigator.language || "en").toLowerCase().startsWith("zh"); let lang = saved || (navZh ? "zh-CN" : "en"); if (!LANGS.some((l) => l.code === lang)) lang = "en"; const subs = new Set(); function t(s, vars) { if (s == null) return s; let str = (lang === "en") ? s : (DICT[lang] && DICT[lang][s] !== undefined ? DICT[lang][s] : s); if (vars) { for (const k in vars) { str = str.replace(new RegExp("\\{" + k + "\\}", "g"), vars[k]); } } return str; } function setLang(code) { if (!LANGS.some((l) => l.code === code) || code === lang) return; lang = code; localStorage.setItem("aim.lang", code); subs.forEach((fn) => fn(code)); } window.AII18N = { LANGS, get lang() { return lang; }, t, setLang, subscribe(fn) { subs.add(fn); return () => subs.delete(fn); }, }; })();