[Web] bug(webgpu): Error: FILTER_IN_CHANNEL should be equal to DATA_CHANNEL #27277
Description
Describe the issue
When attempting to run inference with WebGPU on a specific model, I encounter the error shown in the title.
Since this issue does not occur with the WASM Execution Provider, I believe it is likely specific to WebGPU.
For details, please refer to the reproduction steps.
To reproduce
Save the following code as HTML and open it in browser.
<!doctype html>
<html lang="en">
<head> </head>
<body>
<code id="log" style="white-space: pre-wrap"></code>
<script src="https://cdn.jsdelivr.net/npm/onnxruntime-web@1.24.1/dist/ort.all.min.js"></script>
<script>
let logText = "";
async function runTest(providerName = "webgpu") {
log.innerText = logText += `[${providerName}] initializing...\n`;
try {
ort.env.wasm.wasmPaths =
"https://cdn.jsdelivr.net/npm/onnxruntime-web@1.24.1/dist/";
ort.env.wasm.numThreads = 1;
ort.env.wasm.simd = true;
const onnxUrl =
"https://huggingface.co/kaeru-shigure/sandbox-onnx/resolve/main/mobilenet_vocals.onnx";
const session = await ort.InferenceSession.create(onnxUrl, {
executionProviders: [providerName],
externalData: [
{
path: "mobilenet_vocals.onnx.data",
data: onnxUrl + ".data",
},
],
});
const result = await session.run({
[session.inputMetadata[0].name]: new ort.Tensor(
"float32",
new Float32Array(1 * 2 * 1323000),
[1, 2, 1323000],
),
});
const outputName = Object.keys(result)[0];
const outputTensor = result[outputName];
log.textContent =
logText += `[${providerName}] inference ok: output tensor shape=${outputTensor.dims}\n`;
} catch (error) {
window.lastError = error;
console.error(error);
log.textContent =
logText += `[${providerName}] inference failed: ${error.stack ?? error.message}\n`;
}
}
(async () => {
await runTest("webgpu");
await runTest("wasm");
})();
</script>
</body>
</html>Result
[webgpu] initializing...
[webgpu] inference failed: Error: [WebGPU] Kernel "[Conv] node_Conv_3503_token_566" failed. Error: FILTER_IN_CHANNEL should be equal to DATA_CHANNEL
at Object._OrtRun (https://cdn.jsdelivr.net/npm/onnxruntime-web@1.24.1/dist/ort-wasm-simd-threaded.jsep.mjs:2:446)
at async Ti (https://cdn.jsdelivr.net/npm/onnxruntime-web@1.24.1/dist/ort.all.min.js:4585:26841)
at async aa.run (https://cdn.jsdelivr.net/npm/onnxruntime-web@1.24.1/dist/ort.all.min.js:4585:34287)
at async r.run (https://cdn.jsdelivr.net/npm/onnxruntime-web@1.24.1/dist/ort.all.min.js:6:18533)
at async runTest (converter_debug.html:29:26)
at async converter_debug.html:49:9
[wasm] initializing...
[wasm] inference ok: output tensor shape=1,2,1323000
Urgency
No response
ONNX Runtime Installation
Released Package
ONNX Runtime Version or Commit ID
1.24.1
Execution Provider
'webgpu' (WebGPU)