[WebGPU] Error running Xenova/musicgen-small

[vonner04]

System Info

@huggingface/transformers": "^3.5.1"
"react": "^18.2.0"

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

I am using the musicgen-web example from the example in transformer.js repo.

I updated the imports accordingly to the huggingface/transformers version in system info.

When I change the device from wasm to webgpu the music that gets generated is just a loud electrical hum.

2025-05-04.10-50-35.mp4

Reproduction

Steps to reproduce the behaviour:

1. clone the repository
2. use the musicgen-web example
3. update the dependency via npm install @huggingface/transformers
4. change device wasm to device: webgpu
5. npm install and run the example
6. generate music with any configuration

[xenova]

Hi there 👋 Indeed, I've experienced this issue myself when attempting to upgrade, as there appears to be some precision issues with decoder_model_merged.onnx (only) on WebGPU.

@fs-eire I've been able to reproduce both on JSEP and Native EP.

Example code

import { AutoTokenizer, MusicgenForConditionalGeneration, RawAudio } from '@huggingface/transformers';

// Load tokenizer and model
const tokenizer = await AutoTokenizer.from_pretrained('Xenova/musicgen-small');
const model = await MusicgenForConditionalGeneration.from_pretrained('Xenova/musicgen-small', {
  dtype: {
    text_encoder: 'q4',
    decoder_model_merged: 'q4',
    encodec_decode: 'fp32',
  },
//   device: "webgpu", // <-- uncomment this to enable webgpu (broken)
});

// Prepare text input
const prompt = 'a light and cheerly EDM track, with syncopated drums, aery pads, and strong emotions bpm: 130';
const inputs = tokenizer(prompt);

// Generate audio
const audio_values = await model.generate({
  ...inputs,
  max_new_tokens: 500,
  do_sample: true,
  guidance_scale: 3,
});

// (Optional) Write the output to a WAV file
const audio = new RawAudio(audio_values.data, model.config.audio_encoder.sampling_rate);
audio.save('musicgen.wav');