# ExecuTorch WebGPU Enables Native PyTorch Inference in Web Browsers Run PyTorch models directly on GPU hardware within browsers without format conversion. - **Source:** [PyTorch Blog](https://pytorch.org/blog) - **Published:** 2026-11-10 - **Topic:** [technology](https://brifing.si/c/technology) - **Original article:** - **On Brifing:** https://brifing.si/s/635919ec-a284-481e-9e93-b70a3a9c9137 ## Native PyTorch Web Integration ExecuTorch now supports running PyTorch models directly in web browsers using GPU acceleration. By utilizing the WebGPU backend, developers can execute models without converting them to alternative formats like GGUF or ONNX. This maintains the integrity of the PyTorch toolchain while enabling high-performance inference across major browsers like Chrome and Safari. ## Privacy and Performance Gains Running models locally in the browser ensures that all user data, prompts, and inputs remain on the device. This approach eliminates network latency and reduces server-side infrastructure costs significantly. Benchmarks on an Apple M1 Pro show WebGPU is 4–7x faster on prefill and 2–10x faster on decode compared to CPU-based Wasm execution. ## Unified Model Artifacts The system uses a single .pte file format that functions across diverse hardware platforms. The runtime translates WGSL compute shaders into platform-specific languages like Metal for macOS, Vulkan for Android, or D3D12 for Windows. This architecture ensures consistent performance and deployment without requiring native installations or complex backend server configurations for end users. ## Dynamic Runtime Architecture The backend design relies on three core principles: architecture, export, and dynamic runtime. It handles dynamic shapes through allocate-at-max buffers and per-dispatch workgroup recalculations. This allows the system to manage complex tasks like batched prefill efficiently, while the Emscripten-compiled C++ runtime bridges WebGPU API calls to the browser's native JavaScript interface. ## Broad Model Compatibility The WebGPU backend supports a wide array of model architectures beyond simple text generation. Validated models include LLMs like Llama 3.2 and Gemma 3, vision-language models such as SmolVLM, and specialized tools like Whisper for speech-to-text. This versatility demonstrates that the backend can handle diverse operator requirements across various machine learning domains. ## On-Device Training Capabilities Beyond inference, the framework supports on-device fine-tuning using LoRA and DiReFT techniques. This functionality enables personalized assistants and domain-specific model adaptation directly within the browser sandbox. By keeping training processes local, developers can build adaptive applications that respect user privacy while leveraging the computational power of modern browser-based GPU access. ## Developer Deployment Workflow Deployment is streamlined by serving a single .pte file over a standard CDN. Pipeline compilation, memory management, and dispatch scheduling are handled automatically by the WebGPU backend within the browser tab. This zero-friction approach allows developers to integrate sophisticated AI features into web applications without managing complex native dependencies or servers. ## Future Development Roadmap The ExecuTorch WebGPU backend remains in early development, with ongoing efforts focused on stability and performance. Future updates aim to expand operator coverage, data type support, and overall model compatibility. As an open-source project, the team encourages community contributions, including new operators and performance optimizations, via the official GitHub repository and Discord. ## Sources covering this story - https://pytorch.org/blog/executorch-webgpu-pytorch-models-on-the-gpu-in-the-browser/ — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._