[Web] Deprecate the onnxruntime-web WebGL backend

[edgchen1]

Summary

The legacy WebGL backend in onnxruntime-web is deprecated and planned for removal in a subsequent release. This
issue is the authoritative migration guide and feedback channel for consumers that still rely on WebGL.

Current status

This table will be updated as the deprecation progresses:

Milestone	Status	Relevant change
Phase 1: deprecation warning and migration docs	In progress	PR/commit TBD
Deprecation release	Not released	Version and release notes TBD
Phase 2: WebGL removal	Not scheduled	PR/commit TBD
Removal release	Not scheduled	Version and release notes TBD

Phase 2 is planned for a subsequent release. The schedule may be extended if WebGL-dependent consumers report
migration blockers here.

Phase 1 behavior

The initial deprecation phase does not change inference behavior:

• executionProviders: ['webgl'] continues to select WebGL.
• import 'onnxruntime-web/webgl' and the ort.webgl.* bundles remain available.
• Explicit WebGL use will emit a warning once, subject to ort.env.logLevel.
• WebGL remains opt-in and is not selected as a default fallback.

Why WebGL is being removed

The WebGL backend originated in the standalone ONNX.js project and predates ONNX Runtime's WebAssembly
implementation. It is now in maintenance mode, supports fewer operators than WebGPU and WASM, and has behavioral
and floating-point differences from newer backends. Removing it will reduce duplicate maintenance and simplify the
onnxruntime-web build matrix.

Migration options

WebGPU

Use WebGPU where the target browsers support it:

import * as ort from 'onnxruntime-web/webgpu';

const session = await ort.InferenceSession.create(model, {
  executionProviders: ['webgpu'],
});

WebAssembly/CPU

Use WASM when WebGPU is unavailable or broader browser compatibility is required:

import * as ort from 'onnxruntime-web/wasm';

const session = await ort.InferenceSession.create(model);

The default onnxruntime-web import also uses WASM/CPU when no executionProviders option is specified. The
onnxruntime-web/wasm import explicitly selects the CPU-only bundle.

There will be no automatic redirect from WebGL. The WebGL-only bundle has no WASM fallback, and silently redirecting
to WebGPU would fail on browsers without navigator.gpu and could introduce hidden operator or numerical behavior
changes.

Feedback requested

If you cannot migrate to WebGPU or WASM, please describe:

• the browser/platform that requires WebGL;
• the model or operators involved;
• the WebGPU or WASM failure, compatibility gap, or performance blocker;
• any release-timing constraint we should consider.

Design:
https://github.com/microsoft/onnxruntime/blob/main/docs/design/onnxruntime_web_remove_webgl_backend.md