ExecuTorch adds a WebGPU backend for running PyTorch models in browsers
AIPyTorch says ExecuTorch now runs PyTorch models on the GPU in WebGPU-enabled browsers, using a .pte file exported with torch.export and no conversion to another format. The company reports WebGPU is 4–7× faster than the CPU path (WebAssembly with XNNPACK) on prefill and 2–10× faster on decode, measured in Chrome on an Apple M1 Pro.
AIWhy it matters.
The post shows how a PyTorch model moves from a single export to in-browser GPU inference, with benchmark figures against the existing CPU path.
Loading source posts…
Source details
PyTorch Blog · Julian Ng-Thow-Hing, Siddartha Pothapragada, Stephen Jia, Digant Desai, Bilgin Cagatay · pytorch.orgPublished · added here
Original title: ExecuTorch WebGPU: PyTorch Models on the GPU, in the Browser