ExecuTorch WebGPU Runtime (WebAssembly)
The ExecuTorch runtime compiled to WebAssembly with its WebGPU backend.
It runs ExecuTorch .pte programs delegated to WebGPU inside a web browser, on the visitor's GPU. The WebAssembly code
loads programs and schedules work; the model math runs on the GPU through WebGPU. It powers the ExecuTorch WebGPU
Playground (launching soon on executorch.ai), whose models are in the
ExecuTorch WebGPU collection.
Files
| File | Size | SHA-256 |
|---|---|---|
executorch_wasm_lib.wasm |
6.6 MB | 35a9087d0b6cd86c279e7111adfd54a58156bd993928b20f63348f1043652eb4 |
executorch_wasm_lib.js |
166 KB | 8a631b1537927f8f2ef184d9f1b343061f47df0a32ddb8f5190d74e80d40618e |
The two files are one build and must be used together. executorch_wasm_lib.js defines createExecuTorch(); after
await et.initWebGPU(), a .pte loads with et.Module.load(path) and runs with an async forward([tensors]).
The playground fetches the .wasm from this repository, checks its SHA-256, and passes it to the loader.
Tokenizer library
vendor/transformers/transformers.min.js is @huggingface/transformers 3.0.2, byte-identical to the npm release
(Apache-2.0, vendor/transformers/LICENSE; bundled @huggingface/jinja and onnxruntime-web code under MIT,
vendor/transformers/THIRD_PARTY_NOTICES.txt). The playground uses it only for tokenizers and chat templates; it checks
the file's SHA-256 before importing it.
Build
Built from the public ExecuTorch sources (extension/wasm with EXECUTORCH_BUILD_WASM_WEBGPU=ON) with Emscripten 4.0.19.
BUILD_INFO.md has the toolchain, sources and exact commands.
License
ExecuTorch is BSD-licensed (LICENSE). The build statically includes emdawnwebgpu (Dawn), the Emscripten runtime,
musl, libc++, libc++abi, compiler-rt and FlatBuffers; their licenses are in THIRD_PARTY_NOTICES.txt.