This actually works on my non-M1/M2 macbook with AMD GPUs pretty smoothly. It's a bit slow for now (~70sec), but I expect this to be faster with fp16. Any roadmap to support fp16 on WebGPUs?