GPUAdapter invalidation causes TensorFlow.js WebGPU backend to fail — Follow-up to #22029

[desaisoftwaree]

Version: Deno 2.1.10
(stable, release, x86_64-pc-windows-msvc)
v8 13.0.245.12-rusty
typescript 5.6.2

Library:
TensorFlow.js 4.22.0 (WebGPU backend)

Description: This is a follow-up to the now-closed issue #22029. The core problem persists: using the TensorFlow.js WebGPU backend under Deno leads to adapter invalidation errors.

When TensorFlow.js attempts to initialize the WebGPU backend, it eventually tries to re-check the same GPUAdapter (after a device is created). Deno immediately invalidates a GPUAdapter after the first device creation, causing:
deno run --allow-net=cdn.jsdelivr.net main.js
Initialization of backend webgpu failed
TypeError: The adapter cannot be reused, as a device creation has invalidated it
at GPUAdapter.get info (ext:deno_webgpu/01_webgpu.js:525:13)
at Object.factory (https://cdn.jsdelivr.net/npm/@tensorflow/tfjs-backend-webgpu@4.22.0/+esm:170:749)
at async Vs.ready (https://cdn.jsdelivr.net/npm/@tensorflow/tfjs-core@4.22.0/+esm:186:1575)
at async file:///D:/Dev/deno-web-gpu/main.js:4:1
error: Uncaught (in promise) Error: Could not initialize any backends, all backend initializations failed.
at Vs.ready (https://cdn.jsdelivr.net/npm/@tensorflow/tfjs-core@4.22.0/+esm:186:1658)
at async file:///D:/Dev/deno-web-gpu/main.js:4:1

Steps to Reproduce:

1. Create a simple Deno project with this structure:

deno.json

{
  "importMap": "import_map.json"
}

import_map.json

{
  "imports": {
    "@tensorflow/tfjs-core": "https://cdn.jsdelivr.net/npm/@tensorflow/tfjs-core@4.22.0/+esm",
    "@tensorflow/tfjs-backend-webgpu": "https://cdn.jsdelivr.net/npm/@tensorflow/tfjs-backend-webgpu@4.22.0/+esm"
  }
}

2. main.js

import * as tf from "@tensorflow/tfjs-core";
import "@tensorflow/tfjs-backend-webgpu";

await tf.ready();
tf.randomGamma([2, 2], 1).print();

3. Run:

deno run --allow-net=cdn.jsdelivr.net main.js

4. Observe the error about the adapter not being reusable after device creation.
   Expected Behavior: In browsers (e.g., Chrome, Edge), the WebGPU adapter is not invalidated right after device creation. TensorFlow.js’s WebGPU backend can repeatedly query or reuse the adapter without failing. Ideally, Deno would match the browser’s behaviour closely enough that TF.js does not crash on startup.

Actual Behavior: Deno invalidates the GPUAdapter as soon as a device is created. TensorFlow.js tries to re-check or reuse the adapter, triggering a TypeError: The adapter cannot be reused, as a device creation has invalidated it.

Why This Matters:

• TensorFlow.js (and potentially other libraries) rely on creating or querying the adapter multiple times, mirroring how browsers handle WebGPU.
• Deno’s strict adapter invalidation policy disrupts these libraries, making the WebGPU backend unusable.

Additional Context:

• The originally closed Deno issue is #22029. That issue was closed, but the underlying limitation remains a blocker for TF.js WebGPU usage under Deno.

Possible Approaches:

• Align Deno’s WebGPU adapter lifecycle closer to the browser behaviour so that libraries like TensorFlow.js can function without changes.

• Provide a configuration or override that allows multiple device creations or repeated adapter queries.

Thank you for revisiting this. Looking forward to any updates or clarifications on whether Deno can (or will) change its WebGPU behaviour, or if library developers need to implement a workaround to accommodate Deno’s stricter handling.

[crowlKats]

Fixed in #27665

[desaisoftwaree]

Fixed in #27665

Can you please provide a link or any guide on how to implement it?

[crowlKats]

@desaisoftwaree not sure I understand what you are asking for

[desaisoftwaree]

@desaisoftwaree not sure I understand what you are asking for

I am not able to understand why is this issue closed immediately. It would be great if you can provide any hints.

[petamoriken]

I believe this issue has been already resolved and is not yet released so you can try deno upgrade --canary.

[desaisoftwaree]

I believe this issue has been already resolved and is not yet released so you can try deno upgrade --canary.

I upgraded to canary but the same error persists.

[petamoriken]

Observe the error about the adapter not being reusable after device creation.
Expected Behavior: In browsers (e.g., Chrome, Edge), the WebGPU adapter is not invalidated right after device creation. TensorFlow.js’s WebGPU backend can repeatedly query or reuse the adapter without failing. Ideally, Deno would match the browser’s behaviour closely enough that TF.js does not crash on startup.

My understanding is that there is a divergence between Chrome's implementation and wgpu (used by Firefox and Deno) due to changes in the WebGPU spec. Perhaps a fix on the TensorFlow.js side might be appropriate.
#25874 (comment)

@crowlKats What do you think?

[desaisoftwaree]

You are correct

There is a divergence between Chrome's implementation and wgpu (used by Firefox and Deno) due to changes in the WebGPU spec
I tried running tensorflow js on Firefox and enabled the webgpu from about:config but errors are thrown. Also, WebGPU for Firefox is not yet complete