Run Laya in the browser

Libraries

kevala is a zero-dependency Rust engine compiled to WebAssembly, with WebGPU kernels. It works from any page:

<script type="module">
  import { Kevala } from "https://cdn.jsdelivr.net/npm/kevala@latest/js/src/index.js";
  const kevala = await Kevala.load({ model: "laya", onProgress: console.log });
</script>

or pnpm add kevala. According to its README, the Laya int8 pack is 479 MB and is kept in browser storage after the first visit.

@r4ai/laya-web runs ONNX Runtime Web with WebGPU and a WebAssembly SIMD fallback, and is designed to run in a Web Worker:

npm install @r4ai/laya-web onnxruntime-web

Demos you can open

• layaForWeb: the English checkpoint as quantized ONNX (default build about 440 MB), with a live demo.
• open-jev-laya: multilingual Laya on Transformers.js (fp16 ONNX about 647 MB) with Gomoku, Big Two and a 3D maze.
• layaAsRagJudge: checks RAG claims against retrieved evidence entirely in the tab.

Speed and support

The layaForWeb README reports that a three-question call on the default WASM backend took about 2 to 5 seconds on a 2-core machine. In that project, WebGPU works only with the int4 build, because ONNX Runtime's WebGPU MatMulNBits kernel supports 2- and 4-bit weights. open-jev-laya tries WebGPU and falls back to WebAssembly.

Caveats

Quantized browser builds do not match PyTorch exactly. layaForWeb reports 97.9% top-answer agreement for all three of its quantized builds, with the largest probability gaps in int4. For a real-time loop, laya-pong keeps the model native. Its README gives the reason: the checkpoint wants about 2.4 GB resident in f32, and wasm32 has a 4 GB address space and no Metal.