Ruvector Attention Wasm

@ruvector/attention-wasm

WebAssembly-compiled attention mechanisms providing browser-native and edge-compatible FlashAttention, MultiHeadAttention, CrossAttention, and LinearAttention with SIMD acceleration and zero server dependency.

Quick Reference

Installation

Install: npx @ruvector/attention-wasm@latest See Installation Guide for the full ecosystem.

Key API

WasmFlashAttention

WASM-compiled memory-efficient FlashAttention with IO-awareness.

Node.js:

import { WasmFlashAttention } from '@ruvector/attention-wasm';

const attn = new WasmFlashAttention({
  heads: 8,
  dim: 64,
  blockSize: 256,
});

Browser:

import init, { WasmFlashAttention } from '@ruvector/attention-wasm';

await init(); // Initialize WASM module

const attn = new WasmFlashAttention({
  heads: 8,
  dim: 64,
  blockSize: 256,
});

Constructor Options:

Methods:

WasmMultiHeadAttention

WASM-compiled standard multi-head attention.

import init, { WasmMultiHeadAttention } from '@ruvector/attention-wasm';
await init();

const mha = new WasmMultiHeadAttention({
  modelDim: 512,
  heads: 8,
});

Constructor Options:

Methods:

WasmCrossAttention

WASM-compiled cross-attention between two sequences.

import init, { WasmCrossAttention } from '@ruvector/attention-wasm';
await init();

const cross = new WasmCrossAttention({
  queryDim: 512,
  keyDim: 768,
  heads: 8,
});

Constructor Options:

WasmLinearAttention

WASM-compiled O(n) linear attention approximation.

import init, { WasmLinearAttention } from '@ruvector/attention-wasm';
await init();

const linear = new WasmLinearAttention({
  dim: 64,
  heads: 8,
  featureMap: 'elu',
});

Constructor Options:

Common Patterns

Browser Transformer Inference

import init, { WasmFlashAttention } from '@ruvector/attention-wasm';

await init();

const selfAttn = new WasmFlashAttention({ heads: 8, dim: 64, causal: true });
const output = selfAttn.forward(queryTokens, keyTokens, valueTokens);

// Render output in browser
document.getElementById('result').textContent = decodeOutput(output);

Edge Device Attention

import { WasmLinearAttention } from '@ruvector/attention-wasm';

// O(n) linear attention for resource-constrained devices
const attn = new WasmLinearAttention({ dim: 32, heads: 4, featureMap: 'elu' });
const output = attn.forward(Q, K, V);
console.log(`Memory: ${attn.getMemoryUsage().heapUsed} bytes`);

Bundler Configuration (Vite)

// vite.config.ts
import wasm from 'vite-plugin-wasm';

export default {
  plugins: [wasm()],
  optimizeDeps: { exclude: ['@ruvector/attention-wasm'] },
};

RAN DDD Context

Bounded Context: Learning

References

• API reference: See references/commands.md
• Full README
• npm