// CRITICALimport {  Counter,  Gauge,  Histogram,  Registry,} from "prom-client"; /** * Metrics registry wrapper. */export interface MetricsRegistry {  registry: Registry;  contentType: string;  getMetrics: () => Promise<string>;} /** * Controller metrics collection. */export interface ControllerMetrics {  recordModelSwitch: (recipeId: string, backend: string, durationSeconds: number, success: boolean) => void;  updateActiveModel: (modelPath?: string | null, backend?: string | null, servedName?: string | null) => void;  updateGpuMetrics: (gpus: Record<string, unknown>[]) => void;  updateSseMetrics: (stats: Record<string, unknown>) => void;} /** * Create a Prometheus metrics registry and helpers. * @returns Registry helpers and metric update functions. */export const createMetrics = (): { registry: MetricsRegistry; metrics: ControllerMetrics } => {  const registry = new Registry();   const modelSwitchesTotal = new Counter({    name: "vllm_studio_model_switches_total",    help: "Total number of model switches",    labelNames: ["recipe_id", "backend"],    registers: [registry],  });   const modelSwitchDuration = new Histogram({    name: "vllm_studio_model_switch_duration_seconds",    help: "Time taken to switch models",    labelNames: ["recipe_id"],    buckets: [10, 30, 60, 120, 300, 600],    registers: [registry],  });   const modelLaunchFailures = new Counter({    name: "vllm_studio_model_launch_failures_total",    help: "Total number of failed model launches",    labelNames: ["recipe_id"],    registers: [registry],  });   const activeModelInfo = new Gauge({    name: "vllm_studio_active_model",    help: "Currently active model information",    labelNames: ["model_path", "backend", "served_model_name"],    registers: [registry],  });   const inferenceServerUp = new Gauge({    name: "vllm_studio_inference_server_up",    help: "Whether inference server is running (1=up, 0=down)",    registers: [registry],  });   const gpuMemoryUsed = new Gauge({    name: "vllm_studio_gpu_memory_used_bytes",    help: "GPU memory used in bytes",    labelNames: ["gpu_id", "gpu_name"],    registers: [registry],  });   const gpuMemoryTotal = new Gauge({    name: "vllm_studio_gpu_memory_total_bytes",    help: "Total GPU memory in bytes",    labelNames: ["gpu_id", "gpu_name"],    registers: [registry],  });   const gpuUtilization = new Gauge({    name: "vllm_studio_gpu_utilization_percent",    help: "GPU utilization percentage",    labelNames: ["gpu_id", "gpu_name"],    registers: [registry],  });   const gpuTemperature = new Gauge({    name: "vllm_studio_gpu_temperature_celsius",    help: "GPU temperature in Celsius",    labelNames: ["gpu_id", "gpu_name"],    registers: [registry],  });   const sseActiveConnections = new Gauge({    name: "vllm_studio_sse_active_connections",    help: "Number of active SSE connections",    labelNames: ["channel"],    registers: [registry],  });   const sseEventsPublished = new Counter({    name: "vllm_studio_sse_events_published_total",    help: "Total SSE events published",    labelNames: ["event_type"],    registers: [registry],  });   let lastEventCount = 0;   const metrics: ControllerMetrics = {    recordModelSwitch: (recipeId, backend, durationSeconds, success) => {      if (success) {        modelSwitchesTotal.labels({ recipe_id: recipeId, backend }).inc();        modelSwitchDuration.labels({ recipe_id: recipeId }).observe(durationSeconds);      } else {        modelLaunchFailures.labels({ recipe_id: recipeId }).inc();      }    },    updateActiveModel: (modelPath, backend, servedName) => {      activeModelInfo.reset();      const labels = {        model_path: modelPath ?? "",        backend: backend ?? "",        served_model_name: servedName ?? "",      };      activeModelInfo.labels(labels).set(1);      inferenceServerUp.set(modelPath ? 1 : 0);    },    updateGpuMetrics: (gpus) => {      for (const gpu of gpus) {        const gpuId = String(gpu["id"] ?? gpu["index"] ?? 0);        const gpuName = String(gpu["name"] ?? "Unknown");        const labels = { gpu_id: gpuId, gpu_name: gpuName };        let memoryUsed = Number(gpu["memory_used"] ?? 0);        let memoryTotal = Number(gpu["memory_total"] ?? 0);        if (memoryUsed < 1_000_000) {          memoryUsed = memoryUsed * 1024 * 1024;          memoryTotal = memoryTotal * 1024 * 1024;        }        gpuMemoryUsed.labels(labels).set(memoryUsed);        gpuMemoryTotal.labels(labels).set(memoryTotal);        const utilization = Number(gpu["utilization"] ?? gpu["utilization_pct"] ?? 0);        const temperature = Number(gpu["temperature"] ?? gpu["temp_c"] ?? 0);        gpuUtilization.labels(labels).set(utilization);        gpuTemperature.labels(labels).set(temperature);      }    },    updateSseMetrics: (stats) => {      const channels = stats["channels"];      if (channels && typeof channels === "object") {        for (const [channel, count] of Object.entries(channels)) {          sseActiveConnections.labels({ channel }).set(Number(count));        }      }      const totalEvents = Number(stats["total_events_published"] ?? 0);      if (totalEvents > lastEventCount) {        sseEventsPublished.labels({ event_type: "all" }).inc(totalEvents - lastEventCount);        lastEventCount = totalEvents;      } else if (totalEvents < lastEventCount) {        lastEventCount = totalEvents;      }    },  };   const metricsRegistry: MetricsRegistry = {    registry,    contentType: registry.contentType,    getMetrics: async () => registry.metrics(),  };   return { registry: metricsRegistry, metrics };};