Implementation:Microsoft Onnxruntime Session Run Nodejs

Field	Value
Implementation Name	Session_Run_Nodejs
Overview	Asynchronous execution of ONNX model inference using the session.run() method with named input tensors in Node.js.
Type	API Doc
Language	JavaScript
Domains	ML_Inference, JavaScript_Integration
Source Repository	microsoft/onnxruntime
Last Updated	2026-02-10

Overview

Asynchronous execution of ONNX model inference using the session.run() method with named input tensors in Node.js. This method accepts a feeds dictionary mapping input names to ort.Tensor objects and returns a Promise resolving to a results dictionary mapping output names to output tensors.

API

session.run(feeds: Record<string, ort.Tensor>) -> Promise<Record<string, ort.Tensor>>

Source Code Reference

• Repository: microsoft/onnxruntime
• Primary Source: samples/nodejs/01_basic-usage/index.js:L24

The sample model in the basic usage example contains a single MatMul node with two inputs a (float32, 3x4) and b (float32, 4x3) and one output c (float32, 3x3).

Key Parameters

Parameter	Type	Required	Description
feeds	Record<string, ort.Tensor>	Yes	Object mapping model input names to ort.Tensor objects containing input data

I/O Contract

Direction	Name	Type	Description
Input	feeds	Record<string, ort.Tensor>	Dictionary of named input tensors; keys must match model input names
Output	results	Promise<Record<string, ort.Tensor>>	Promise resolving to a dictionary of named output tensors

Usage Examples

Basic MatMul Inference

const ort = require('onnxruntime-node');

async function main() {
    // Create session from model
    const session = await ort.InferenceSession.create('./model.onnx');

    // Prepare input tensors
    const dataA = Float32Array.from([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]);
    const dataB = Float32Array.from([10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120]);
    const tensorA = new ort.Tensor('float32', dataA, [3, 4]);
    const tensorB = new ort.Tensor('float32', dataB, [4, 3]);

    // Build feeds object with model input names as keys
    const feeds = { a: tensorA, b: tensorB };

    // Run inference
    const results = await session.run(feeds);

    // Access output tensor
    const dataC = results.c.data;
    console.log(`data of result tensor 'c': ${dataC}`);
}
main();

Inspecting Output Metadata

const results = await session.run(feeds);

// Each output is an ort.Tensor with type, dims, data, and size
for (const [name, tensor] of Object.entries(results)) {
    console.log(`Output '${name}': type=${tensor.type}, dims=${tensor.dims}, size=${tensor.size}`);
}

Key Details

• The feeds object keys must exactly match the input names defined in the ONNX model graph.
• The method is asynchronous; inference computation happens on native background threads via N-API.
• Input tensor shapes and types must match the model's expected input specifications.
• The returned results object maps output names to ort.Tensor instances containing the computed values.
• Each call to run() executes one complete forward pass through the model graph.
• The Node.js event loop remains responsive during inference computation.

Related Pages

• Principle:Microsoft_Onnxruntime_Nodejs_Inference_Execution
• Implementation:Microsoft_Onnxruntime_InferenceSession_Create_Nodejs
• Implementation:Microsoft_Onnxruntime_Ort_Tensor_Nodejs
• Implementation:Microsoft_Onnxruntime_Tensor_Data_Extraction_Nodejs
• Environment:Microsoft_Onnxruntime_Nodejs_Runtime_Environment