Model Management

Utility functions for fetching, caching, and initializing onnxruntime-web inference sessions.

import {
  loadModels,
  fetchModel,
  releaseModels,
  isCrossOriginIsolated,
  type ModelUrls,
  type LoadOptions,
  type LoadedModels,
} from 'vigilo-wasm';

---

Functions

loadModels(urls: ModelUrls, options?: LoadOptions): Promise<LoadedModels>

Downloads .onnx model graphs (using the Cache API when available) and creates InferenceSession instances for each slot.

const models = await loadModels({
  face:    '/models/face_detection_yunet_2023mar.onnx',
  pose:    '/models/headpose_mobilenetv3_small.onnx',
  gaze:    '/models/mobileone_s0_gaze.onnx',
  objects: '/models/yolox_nano.onnx',
}, {
  executionProviders: ['webgpu', 'wasm'],
  cache: true,
  onProgress: (slot, loaded, total) => console.log(`${slot}: ${loaded}/${total}`),
});

ModelUrls

• face: string (Required): YuNet face detection model.
• pose?: string (Optional): MobileNetV3 head pose model.
• gaze?: string (Optional): MobileOne-S0 gaze tracking model.
• objects?: string (Optional): YOLOX-Nano prohibited object model.

LoadOptions

• ort?: OrtLike: onnxruntime-web namespace override.
• executionProviders?: string[]: Execution provider priority, e.g. ['webgpu', 'wasm'].
• wasmPaths?: string: Path or URL where ORT's .wasm binaries are hosted.
• numThreads?: number: Thread count for WASM provider (defaults to navigator.hardwareConcurrency if cross-origin isolated, else 1).
• cache?: boolean: Whether to cache model weights in the Cache API (default true).
• onProgress?: (slot: ModelSlot, loaded: number, total: number) => void: Download progress callback.

LoadedModels

interface LoadedModels {
  face: InferenceSession;
  pose?: InferenceSession;
  gaze?: InferenceSession;
  objects?: InferenceSession;
  failed: Array<{ slot: ModelSlot; error: string }>;
}

---

fetchModel(url: string, slot: ModelSlot, options?: { cache?: boolean; onProgress?: ... }): Promise<Uint8Array>

Fetches raw bytes of an .onnx model file, checking the Cache API bucket vigilo-models-v1 first.

---

releaseModels(models: LoadedModels): Promise<void>

Releases all underlying WebAssembly / WebGPU session memory in onnxruntime-web.

---

isCrossOriginIsolated(): boolean

Returns true if globalThis.crossOriginIsolated is enabled, indicating that SharedArrayBuffer is available for multi-threaded WASM inference.