@agentix-e/timesfm-node

Node.js ONNX Runtime inference engine for TimesFM — zero-shot time series forecasting.

📚 API Documentation · 📊 Benchmark · 📈 Coverage · 💻 Source

Overview

@agentix-e/timesfm-node provides the default Node.js inference engine for TimesFM. It implements the IInferenceEngine interface from @agentix-e/timesfm-core using onnxruntime-node.

This package is automatically loaded by TimesFMModel.fromPretrained() when no custom engine is provided. Browser users who inject TimesFMWebInferenceEngine (from @agentix-e/timesfm-web) never install this package or the native ONNX Runtime addon.

npm install @agentix-e/timesfm-node

Usage

import { TimesFMModel, createForecastConfig } from '@agentix-e/timesfm-core';
// The model automatically loads timesfm-node when no engine is provided
const model = await TimesFMModel.fromPretrained({
  modelPath: './timesfm-2.5.onnx',
});

Explicit engine creation:

import { TimesFMNodeEngine, createDefaultEngine } from '@agentix-e/timesfm-node';

// Option A: Direct construction
const engine = new TimesFMNodeEngine(TIMESFM_25_CONFIG, {
  executionProvider: 'cpu', // 'cpu' | 'cuda' | 'dml'
  intraOpNumThreads: 4,
});

// Option B: Factory (used internally by TimesFMModel)
const engine = createDefaultEngine(config, { executionProvider: 'cpu' });

const model = await TimesFMModel.fromPretrained({
  modelPath: './timesfm-2.5.onnx',
  engine,
});

Exports

| Export | Description | | --------------------- | -------------------------------------------------------------- | | TimesFMNodeEngine | IInferenceEngine implementation backed by onnxruntime-node | | createDefaultEngine | Factory function for creating the default engine |

Execution Providers

| Provider | Description | | -------- | ---------------------------------------------------------- | | cpu | CPUExecutionProvider (default) | | cuda | CUDAExecutionProvider (requires NVIDIA GPU + CUDA drivers) | | dml | DmlExecutionProvider (requires DirectML on Windows) |

System Requirements

• Node.js ≥ 22
• 4+ GB RAM
• 2 GB disk space for the ONNX model (~885 MB zipped, ~928 MB extracted)
• CUDA 12+ (optional, for GPU inference)