Error: using ceil() in shape computation is not yet supported for MaxPool

[geronimi73]

System Info

@huggingface/transformers 3.7.5
Next.js 14.2.2
Chrome, M3 Macbook

Environment/Platform

• Website/web-app
• Browser extension
• Server-side (e.g., Node.js, Deno, Bun)
• Desktop app (e.g., Electron)
• Other (e.g., VSCode extension)

Description

Model: onnx-community/ormbg-ONNX

Processing fails with error (see bottom) if model loaded with webgpu device (Chrome, Mac M3).

this.model = await pipeline(
    'background-removal', 
    'onnx-community/ormbg-ONNX',
    {
        device: "webgpu",
        dtype: 'fp16',
    }
);

Works beautiful if loaded on cpu:

this.model = await pipeline(
    'background-removal', 
    'onnx-community/ormbg-ONNX',
    {
        device: "wasm",
        dtype: 'fp16',
    }
);

Error:

transformers.web.js:8629 An error occurred during model execution: "Error: using ceil() in shape computation is not yet supported for MaxPool".
sessionRun	@	transformers.web.js:8629

Inputs given to model: 
Object pixel_values:  {type: 'float32', dims: Array(4), location: 'cpu', data: Float32Array(3145728)}
[[Prototype]]: Object 
sessionRun	@	transformers.web.js:8630

Uncaught (in promise) Error: using ceil() in shape computation is not yet supported for MaxPool
    at Array.bl (ort.bundle.min.mjs?d796:2252:627)
    at Co.computeKernel (ort.bundle.min.mjs?d796:2838:10416)
    at Object.eval [as ac] (ort.bundle.min.mjs?d796:2838:17635)
    at 845314 (ort-wasm-simd-threaded.jsep.mjs:48:297)
    at ic (ort-wasm-simd-threaded.jsep.mjs:94:443)
    at ort-wasm-simd-threaded.jsep.wasm:0x103a1a6
    at ort-wasm-simd-threaded.jsep.wasm:0x103ae8c
    at ort-wasm-simd-threaded.jsep.wasm:0xf3c61b
    at ort-wasm-simd-threaded.jsep.wasm:0x261d37
    at ort-wasm-simd-threaded.jsep.wasm:0x706404

Reproduction

const model = await pipeline(
    'background-removal', 
    'onnx-community/ormbg-ONNX',
    {
        device: "webgpu",
        dtype: 'fp16',
    }
);
const rawImage = (await model( rawImage ))[0];
rawImage.toCanvas() 

// .. display canvas

[xenova]

Hi there 👋 This will be fixed when we upgrade to the new WebGPU EP in #1382. Until then, you can use a different model which does work with WebGPU, like:

import { pipeline } from "https://cdn.jsdelivr.net/npm/@huggingface/transformers@3.7.5";
const segmenter = await pipeline("background-removal", "onnx-community/BEN2-ONNX", {
    device: "webgpu",
});
const output = await segmenter("https://images.unsplash.com/photo-1503342217505-b0a15ec3261c?q=80&w=2070&auto=format&fit=crop");
output[0].save("output.png"); // (Optional) Save the image

Example output:

Before	After