Quick question to the Fediverse: has anyone successfully implemented image embedding in the browser *with WebGPU support*?

All my attempts seem to fail. Regardless of wether im using Transformers.js or onnxruntime-web, pre-built models or ONNX files exported myself – at runtime, processing always seems to fall back to CPU.

(Interestingly: segmentation with a self-exported FastSAM model runs fine and *does* use WebGPU.)