[Documentation] WebGPU docs show `Conv` supports `11+`, but JS EP only matches opset 11 (Conv‑22 fails)

[sharwell]

Hi team — I ran into a mismatch between WebGPU operator docs and actual kernel matching for Conv in onnxruntime‑web.

What I observed

• The generated docs js/web/docs/webgpu-operators.md show:

  onnxruntime/js/web/docs/webgpu-operators.md

  Line 32 in 25a6fdc

  | Conv | ai.onnx(1-10,11+); com.ms.internal.nhwc(1-10,11+) | need perf optimization; conv3d is not supported; need implementing activation |

• But in practice, Conv‑22 nodes fail on WebGPU with “kernel not found” (falls back to CPU execution).

• Re‑exporting the model to opset 21 makes it run (uses Conv-11 which is the version preceding Conv-22).

Why this happens

The WebGPU docs are generated by js/web/script/generate-webgpu-operator-md.ts by parsing kernel registrations in:

• onnxruntime/core/providers/js/js_execution_provider.cc
• onnxruntime/core/providers/js/operators/conv.cc

In those files, Conv is registered as:

• 1-10 via ONNX_OPERATOR_VERSIONED_KERNEL_EX
• 11 via ONNX_OPERATOR_KERNEL_EX

In ORT kernel matching, ONNX_OPERATOR_KERNEL_EX(..., 11, ...) sets SinceVersion(11) without an end version (leaving the default value INT_MAX), and kernel matching only treats it as exactly 11, not “11+”. This is enforced in onnxruntime/core/framework/kernel_registry.cc:

onnxruntime/onnxruntime/core/framework/kernel_registry.cc

Lines 126 to 133 in 25a6fdc

	bool valid_version =
	// exact match. typical usage.
	kernel_start_version == since_ver ||
	// allow match if the kernel def has an end version. if it does not, all we know is that the kernel supported
	// the start version when it was created, and not whether a new version of the operator was added since then
	// that the kernel doesn't support.
	(kernel_end_version != INT_MAX &&
	kernel_start_version <= since_ver && kernel_end_version >= since_ver);

So Conv‑22 does not match the “11” kernel and fails.

Impact

• Docs say 11+, but the WebGPU EP rejects Conv‑22.
• This is confusing for users and leads to unexpected runtime failures.

Suggested fixes

Either:

1. Update the docs generator to avoid showing 11+ for ONNX_OPERATOR_KERNEL_CLASS_NAME(..., 11, ...) unless a version range is registered, or
2. Register Conv explicitly for 11–21 and 22 (matching native WebGPU EP style), so the docs become true.

Repro

• Any model with opset import 22 and a Conv node fails on onnxruntime-web WebGPU.
• The same model converted to opset 11 runs.

I’m happy to submit a PR if desired — just confirm the preferred direction (doc fix vs. kernel registration).

Thanks!

Page / URL

No response