ExecuTorch Brings PyTorch Models to Browser GPUs with WebGPU

Introduction

Running a PyTorch model in a browser has often meant converting it into another runtime’s format before dealing with browser-specific execution and device compatibility. ExecuTorch’s new WebGPU backend takes a different route. Developers export a model with torch.export into an ExecuTorch .pte file, then load that artifact in a WebGPU-capable browser through the ExecuTorch runtime.

What matters

• A more direct PyTorch path. The WebGPUPartitioner assigns supported operations to the WebGPU backend, where WGSL compute shaders perform the work. Developers do not need to make ONNX, GGUF, or TFLite the primary deployment format.
• GPU execution inside the tab. The dynamic C++ runtime is compiled to WebAssembly with Emscripten. The emdawnwebgpu layer connects it to the browser’s WebGPU JavaScript API. The same shader path can map to Metal on macOS, Vulkan on Linux and Android, and D3D12 on Windows.
• More than language models. Validation includes Llama, Qwen, and Gemma models, as well as SmolVLM, depth estimation, SAM 2 segmentation, object detection, super-resolution, Whisper speech recognition, and text-to-speech. The project also demonstrates LoRA and DiReFT fine-tuning for selected language models on the device.
• A meaningful CPU comparison. In Chrome on an Apple M1 Pro, quantized Llama 3.2 1B and Qwen3-0.6B were compared with ExecuTorch’s multithreaded Wasm/XNNPACK path. WebGPU was about 4–7 times faster for prefill and 2–10 times faster for decode in that setup. These figures describe the tested hardware and browser, rather than a universal guarantee.

Architecture and caveats

The runtime handles dynamic shapes with buffers allocated to a maximum size, recalculating workgroups for each dispatch and routing shapes to specialized WGSL kernels. This supports batched prefill while allowing one .pte artifact to target different browsers and devices. Still, the backend is explicitly at an early stage. Operator and data-type coverage, stability, and performance all remain active areas of development.

Why it matters

The browser is one of the widest software distribution channels. Local inference can reduce network latency, support offline-oriented experiences, and lower the need for server-side inference. It also keeps prompts, images, camera frames, and audio within the device, which is valuable for privacy-sensitive assistants and media tools.

There are practical limits. GPU capability, WebGPU availability, model size, initial download time, and browser memory behavior will shape the user experience. ExecuTorch WebGPU is therefore best viewed as a promising deployment path rather than proof that every PyTorch model is ready for the browser. Its long-term usefulness will depend on broader operator support and consistent performance across real devices.

The ExecuTorch WebGPU Playground provides a hands-on demo, while model repositories include .pte artifacts and export recipes.

PyTorch Blog