Class TimesFMWebInferenceEngine

Browser-compatible inference engine backed by onnxruntime-web.

Supports WASM, WebGPU, and WebGL backends with automatic fallback. Accepts model URLs (fetched) or pre-loaded ArrayBuffers.

Implements

IInferenceEngine

Index

Constructors

Methods

setWasmPath load forward dispose isLoaded

Constructors

constructor

new TimesFMWebInferenceEngine(
    config: ModelConfig,
    executionProviders?: ("webgpu" | "wasm" | "webgl")[],
    cdnVersion?: string,
    logger?: WebEngineLogger,
): TimesFMWebInferenceEngine
Parameters
- config: ModelConfig
  Model architecture configuration.
- executionProviders: ("webgpu" | "wasm" | "webgl")[] = ...
  Execution providers to try, in order. Default: ['webgpu', 'wasm']
- cdnVersion: string = '1.22.0'
  onnxruntime-web CDN version for browser WASM fallback. Default: matches the package's peerDependency version.
- Optionallogger: WebEngineLogger
  Optional structured logger for diagnostics. When omitted, provider selection messages are silenced (only the final loaded provider is logged).
Returns TimesFMWebInferenceEngine
- Defined in packages/timesfm-web/src/web-engine.ts:126

Methods

setWasmPath

setWasmPath(wasmPath: string): void
Set a custom WASM binary path for onnxruntime-web.

Required when running in Node.js (not a browser) since the default CDN URL won't work. Point this to the dist/ directory of the onnxruntime-web package.
Parameters
- wasmPath: string
Returns void
Example
```
// In Node.js testing:
import { createRequire } from 'node:module';
const require = createRequire(import.meta.url);
const wasmDir = require.resolve('onnxruntime-web').replace('/lib/index.js', '/dist/');
engine.setWasmPath(wasmDir);
```
- Defined in packages/timesfm-web/src/web-engine.ts:154

load

load(
modelPath: string | ArrayBuffer,
options?: { skipWarmup?: boolean },
): Promise<void>
Load the ONNX model.
Parameters
- modelPath: string | ArrayBuffer
  URL to the ONNX model file, or an ArrayBuffer.
- Optionaloptions: { skipWarmup?: boolean }
Returns Promise<void>
Implementation of IInferenceEngine.load
- Defined in packages/timesfm-web/src/web-engine.ts:167

forward

forward(
inputs: Float32Array<ArrayBufferLike>[],
masks: Uint8Array<ArrayBufferLike>[],
): Promise<RawModelOutput>
Run inference for a batch of patched time series.

Aligned with TimesFMNodeEngine.forward: each batch element is processed sequentially via the ONNX session and results are returned as per-element arrays in the RawModelOutput.

Builds the tokenizer input identically to the Node.js engine: each 64-dim patch is [patch_values(32), patch_mask(32)] where mask=0 means "visible" and mask=1 means "padding/ignored".
Parameters
- inputs: Float32Array<ArrayBufferLike>[]
- masks: Uint8Array<ArrayBufferLike>[]
Returns Promise<RawModelOutput>
Implementation of IInferenceEngine.forward
- Defined in packages/timesfm-web/src/web-engine.ts:267

dispose

dispose(): Promise<void>
Release all resources.

Returns Promise<void>
Implementation of IInferenceEngine.dispose
- Defined in packages/timesfm-web/src/web-engine.ts:402

isLoaded

isLoaded(): boolean
Whether the engine is loaded and ready.

Returns boolean
Implementation of IInferenceEngine.isLoaded
- Defined in packages/timesfm-web/src/web-engine.ts:422