PyTorch launches ExecuTorch WebGPU GPU inference

PyTorch models can now run directly in web browsers on the GPU using ExecuTorch without converting to another framework's format. Exported .pte model files operate through a new WebGPU backend that leverages WebAssembly and WGSL compute shaders, achieving performance 4 to 10 times faster than ExecuTorch's previous CPU path. The backend supports a broad range of local AI architectures, including large language models, vision-language systems, depth estimation, and speech-to-text. Open article →