[WebGPU/js] `failed to inference ONNX model: Error: [WebGPU] Kernel "[Add] /decoder/F0.1/Add" failed. Error: Can't perform binary op on the given tensors.` - Kokoro TTS

[xenova]

Describe the issue

Unable to run https://huggingface.co/onnx-community/Kokoro-82M-ONNX on WebGPU

2025-01-17 02:57:22.385599 [E:onnxruntime:, sequential_executor.cc:505 ExecuteKernel] Non-zero status code returned while running Add node. Name:'/decoder/F0.1/Add' Status Message: Failed to run JSEP kernel

The nodes are two float32 tensors, so I don't see why this should cause an issue.

cc @guschmue

To reproduce

https://jsfiddle.net/hj9wfn36/

Urgency

Blocks WebGPU support for Transformers.js and kokoro-js

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.21.0-dev.20250114-228dd16893

Execution Provider

'webgpu' (WebGPU)

[maifeeulasad]

Input shape:

• float32[unk__413, 256, unk__415]
• float32[unk__416, 256, unk__418]

I think the size is incompatible for Add operation.

But it's weird when it's working CPU but failing in GPU.

[guschmue]

yes, one of the input shapes for add (coming from the conv) is 1 off.

inputA [1,269,256]
inputB [1,270,256]