microsoft / onnxruntime

ONNX Runtime: cross-platform, high performance ML inferencing and training accelerator
https://onnxruntime.ai
MIT License
13.51k stars 2.76k forks source link

[Web] `Error: [WebGPU] Kernel "[Conv] /text_encoder/encoder/layers.0/feed_forward/conv_2/Conv" failed. Error: FILTER_IN_CHANNEL should be equal to DATA_CHANNEL` #21108

Open xenova opened 2 weeks ago

xenova commented 2 weeks ago

Describe the issue

Attempting to run Xenova/mms-tts-eng on WebGPU produces the following error:

Uncaught Error: [WebGPU] Kernel "[Conv] /text_encoder/encoder/layers.0/feed_forward/conv_2/Conv" failed. Error: FILTER_IN_CHANNEL should be equal to DATA_CHANNEL

To reproduce

  1. Download the model from the HF hub: https://huggingface.co/Xenova/mms-tts-eng/resolve/main/onnx/model.onnx?download=true
  2. Load the session and set execution provider to webgpu
  3. Run the model with attention_mask and input_ids as inputs
  4. Observe error.

Note that the WASM backend does not have this issue (the model runs correctly)

Urgency

Blocks vits/mms model usage in Transformers.js

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.18.0

Execution Provider

'webgpu' (WebGPU)

xenova commented 2 weeks ago

cc @guschmue