The time it takes to execute setBackend varies across different backends.

[jhl13]

import * as tf from '@tensorflow/tfjs';
import { model } from '@tensorflow/tfjs';
import { setWasmPaths, setThreadsCount, getThreadsCount } from '@tensorflow/tfjs-backend-wasm';
setWasmPaths('https://cdn.jsdelivr.net/npm/@tensorflow/tfjs-backend-wasm/wasm-out/');
import '@tensorflow/tfjs-backend-webgpu';

async function Init() {
    await tf.setBackend("webgpu");
    console.log("done");
}
Init();

Is it normal for the webgpu backend to take a long time when using setBackend? Is there any way to reduce the duration of this operation?

cpu backend

wasm backend

webgl backend

webgpu backend

[qjia7]

Do you switch from another backend? Or it's also very slow even if we set backend to webgpu at first (I mean no backend switch)?

[jhl13]

Do you switch from another backend? Or it's also very slow even if we set backend to webgpu at first (I mean no backend switch)?

No backend switching. Just setting backend to webgpu at first.

[qjia7]

It looks like half of the time is spent in the requestAdapter (~35 ms on my device).

requestAdapter is a webgpu API. We can't do much in TFJS. Maybe we need to file an issue to chromium to see whether this API can be further optimized there. @kainino0x @Jiawei-Shao Is it possible to reduce the time of requestadapter?