[V4] pipeline.dispose() destroys WebGPU device, can't be recreated afterwards

[nico-martin]

System Info

Browser and all environments that support WebGPU.

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

Due to a bug in onnxruntime, after we dispose a pipeline or a model the WebGPU is destroyed, but the JavaScript reference in env.webgpu.device is never cleared, leaving a stale reference to a destroyed device.

This means next time it tries create a WebGOU session, it tries to access the destroyed WebGPU device and fails.

Reproduction

1. install Transformers.js v4 npm i @huggingface/transformers@next

2. create a pipeline with webgpu:

const pipe = await pipeline(
  "feature-extraction",
  "onnx-community/all-MiniLM-L6-v2-ONNX",
  {
    device: "webgpu",
  }
);

3. destroy the pipeline

pipe.dispose();

4. create a new pipeline with webgpu:

const pipe2 = await pipeline(
  "feature-extraction",
  "onnx-community/all-MiniLM-L6-v2-ONNX",
  {
    device: "webgpu",
  }
);

[xenova]

will be fixed when we release a version with ort-web that includes microsoft/onnxruntime#27634 🙌