[Bug] [Node.js binding] Memory leak after releasing inference session

[EliasIBM]

Describe the issue

There's a memory leak occurring when creating an inference session using the runtime in Node. Each time a new session is created and released, the memory usage grows a little. We're running inference on an edge device with limited memory available, so after a couple sessions, the device runs out of memory. We need to reload the session, without restarting the whole application, as it's a realtime system.

Despite disabling "CpuMemArena" and "memPattern", as suggested for similar issues, the memory leak is still occurring.

To reproduce

I've made a minimal javascript example that reproduces the issue:

import * as ort from "onnxruntime-node";

const MODEL_FILE_PATH = "./model.onnx";

const getModel = async (filename) => await ort.InferenceSession.create(filename, { enableCpuMemArena: false, enableMemPattern: false });

const loadSession = async () => {
    const session = await getModel(MODEL_FILE_PATH);
    await session.release();
}

const main = async () => {
    for (let i = 0; i < 10; i++) {
        await loadSession();
    }
}

(async () => {
    await main()
})();

When testing locally, the application memory usage sits at 325 MB when loading the session once. When loading and releasing the session 10 times, it bloats to 994 MB. For 100 load/release cycles it sits at 9.12 GB

Urgency

MEDIUM

Platform

Mac

OS Version

14.5

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.22.0-rev

ONNX Runtime API

JavaScript

Architecture

X64

Execution Provider

Default CPU

Execution Provider Library Version

No response

Model File

No response

Is this a quantized model?

Unknown