We removed wgpu-py dependency from @__tinygrad__ and use Dawn engine through an autogenerated python interface. No need to have another extra python layer when we can just use it directly.

Right?? Either a new season or a new show. We need it back, better than ever.

Thanks @PlisSergey , i modified your cli repo slightly to add support for webgpu export. export_model is not yet in core tinygrad so i used the cloned repo and not pip installed. It’s running in f32 for now. I’ll publish the port repo once done, but you can try out what we have in niivue-tinygrad (tissue_fast variant)

@NTmoney The way he talks about this “thermodynamics quantum ai” garbage is next level cringe.

A cool lil project I'm tackling: tinygrad WebGPU performing brain segmentation in a browser. #tinygrad #webgpu Visualization by niivue. PR: Checkout that performs the model export.

To get nanosecond precision timestamps in Dawn #webgpu engine, you have to enable “allow_unsafe_apis” and disable “timestamp_quantization”. This fixed tinygrad test_search:

dawn also supports f16 datatype, which is a big advantage over wgpu

python bindings to the Dawn (Chrome's) WebGPU engine:

RT @__tinygrad__: Try YOLOv8 on your phone in tinygrad WebGPU! This is it running in Chrome on Z Fold5. You'll need to enable WebGPU on iP…

RT @__tinygrad__: We are going to start putting serious effort into the WebGPU backend. We already linked the Stable Diffusion example, goi…

RT @__tinygrad__: WebGPU is back in tinygrad thanks to @wpmed92 Try Stable Diffusion in Chrome!

Slides and recording are available for the presentation I held at the weekly Open MLIR Meeting. I talked about shaderpulse, a new glsl frontend I’m building. #mlir #llvm #glsl