Show HN: I got PyTorch models running on WebGPU without ONNX export (github.com)

A new experimental WebGPU backend for PyTorch has been announced, providing developers with the ability to run PyTorch code directly on WebGPU devices. Although still in its early stages, the project aims to facilitate high-performance ML model execution without reliance on platform-specific kernels like CUDA and MPS. It achieves this by utilizing a combination of PyTorch, Python, C++, WGSL shaders, and the Google Dawn WebGPU runtime. For example, users can now create and manipulate tensors on the WebGPU device, showcasing seamless data movement between the CPU and WebGPU. This development is significant for the AI/ML community as it opens new avenues for deploying machine learning models across web platforms with greater efficiency. With all major browsers expected to support WebGPU by late 2025, the integration of PyTorch with this technology could enhance the accessibility and scalability of machine learning applications in the web ecosystem. Moreover, the project is positioned for eventual production readiness, with the potential to become part of the core PyTorch offering, thus broadening the toolkit available for developers interested in harnessing GPU capabilities directly within web environments.