
We replaced the wgpu-py WebGPU runtime with the Dawn WebGPU engine (used in Google Chrome) in tinygrad. By using the clang2py tool to auto-generate the Python interface, we directly interact with Dawn, bypassing third-party libraries. This approach allows us to immediately leverage new features from Dawn without waiting for wrapper updates. The integration of Dawn enabled us to finally add f16 support, which was not possible with wgpu.
The next chapter? We’re working on making it incredibly easy to export models to WebGPU. Stay tuned!