index.html

<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<title>Gemma 26B A4B — Browser WebGPU via wllama</title>
<style>
body { font-family: monospace; background: #0d1117; color: #c9d1d9; padding: 24px; max-width: 900px; margin: 0 auto; }
h1 { color: #58a6ff; font-size: 20px; }
.card { background: #161b22; border: 1px solid #30363d; border-radius: 8px; padding: 16px; margin: 12px 0; }
.label { color: #8b949e; font-size: 12px; text-transform: uppercase; letter-spacing: 1px; }
.value { color: #c9d1d9; font-size: 14px; margin-top: 4px; }
.green { color: #3fb950; } .red { color: #f85149; } .amber { color: #d29922; }
#log { font-size: 12px; background: #010409; border: 1px solid #30363d; border-radius: 6px; padding: 10px; max-height: 400px; overflow-y: auto; white-space: pre-wrap; }
button { background: #238636; color: white; border: none; border-radius: 6px; padding: 8px 16px; cursor: pointer; font-weight: bold; margin: 4px; }
button:disabled { opacity: 0.5; cursor: wait; }
input { background: #161b22; border: 1px solid #30363d; color: #c9d1d9; border-radius: 6px; padding: 8px 12px; width: 60%; }
#output { background: #161b22; border: 1px solid #30363d; border-radius: 6px; padding: 12px; min-height: 60px; white-space: pre-wrap; font-size: 14px; margin-top: 8px; }
</style>
</head>
<body>
<h1>Gemma 4 26B A4B — Browser WebGPU</h1>
<p>Gemma-4-26B-A4B-it (MoE, 3.8B active) running in browser via wllama + WebGPU. GGUF loaded from local server.</p>

<div class="card">
  <div class="label">Status</div>
  <div class="value" id="status"><span class="amber">*</span> not initialized</div>
</div>

<div class="card">
  <button id="btn-load" onclick="doLoad()">1. Load Model (WebGPU)</button>
  <button id="btn-gen" onclick="doGenerate()" disabled>2. Generate</button>
</div>

<div class="card">
  <div class="label">Prompt</div>
  <input id="prompt" value="Hello, I am a helpful assistant and" />
</div>

<div class="card">
  <div class="label">Output</div>
  <div id="output"></div>
</div>

<div class="card">
  <div class="label">Log</div>
  <div id="log"></div>
</div>

<script type="module">
import { Wllama } from './node_modules/@wllama/wllama/esm/index.js';

const log = document.getElementById('log');
const status = document.getElementById('status');
const output = document.getElementById('output');
let wllama = null;

function l(msg) {
  const ts = new Date().toISOString().slice(11, 19);
  log.textContent += `[${ts}] ${msg}\n`;
  log.scrollTop = log.scrollHeight;
}

window.doLoad = async function() {
  try {
    document.getElementById('btn-load').disabled = true;
    l('Initializing wllama...');
    status.innerHTML = '<span class="amber">*</span> initializing...';

    const CONFIG_PATHS = {
      default: './node_modules/@wllama/wllama/esm/wasm/wllama.wasm',
    };

    wllama = new Wllama(CONFIG_PATHS, {
      parallelDownloads: 5,
      logger: {
        debug: (msg) => console.log('[wllama]', msg),
        log: (msg) => { console.log('[wllama]', msg); l(msg); },
        warn: (msg) => { console.warn('[wllama]', msg); l('WARN: ' + msg); },
        error: (msg) => { console.error('[wllama]', msg); l('ERROR: ' + msg); },
      },
    });

    l('Loading Gemma 26B A4B (Q5_K_XL, ~20GB in 512MB splits)...');
    l('This will take several minutes on first load.');
    status.innerHTML = '<span class="amber">*</span> loading model...';

    // Load from local server (split GGUF files)
    // wllama auto-detects split pattern from the first file name
    const firstSplit = window.location.origin + '/model/gemma-26b-00001-of-00062.gguf';

    await wllama.loadModelFromUrl(firstSplit, {
      n_gpu_layers: 99, // GPU — patched GLU shader fixes aliasing
      n_ctx: 512, // minimal context to reduce CPU memory
      n_batch: 64,
      useCache: false, // don't cache 20GB in browser storage
      progressCallback: ({ loaded, total }) => {
        const pct = Math.round((loaded / total) * 100);
        if (pct % 5 === 0) l(`Downloading... ${pct}% (${(loaded/1024/1024/1024).toFixed(1)}/${(total/1024/1024/1024).toFixed(1)} GB)`);
        status.innerHTML = `<span class="amber">*</span> downloading ${pct}%...`;
      },
    });

    l('Model loaded!');
    status.innerHTML = '<span class="green">*</span> model ready';
    document.getElementById('btn-gen').disabled = false;
  } catch (e) {
    l('ERROR: ' + e.message);
    console.error(e);
    status.innerHTML = '<span class="red">*</span> ' + e.message;
    document.getElementById('btn-load').disabled = false;
  }
};

window.doGenerate = async function() {
  const prompt = document.getElementById('prompt').value;
  document.getElementById('btn-gen').disabled = true;
  output.textContent = '';
  l('Generating: "' + prompt + '"');
  status.innerHTML = '<span class="amber">*</span> generating...';

  const t0 = performance.now();
  try {
    const result = await wllama.createChatCompletion({
      messages: [{ role: 'user', content: prompt }],
      max_tokens: 500,
      temperature: 0.7,
      top_k: 40,
      top_p: 0.9,
    });

    const elapsed = ((performance.now() - t0) / 1000).toFixed(1);
    console.log('[gemma] raw result:', JSON.stringify(result, null, 2));
    const msg = result?.choices?.[0]?.message;
    const text = msg?.content || '';
    const thinking = msg?.reasoning_content || '';
    const tps = result?.timings?.predicted_per_second?.toFixed(1) || '?';
    if (thinking && !text) {
      output.textContent = thinking;
      l(`[thinking only, ${tps} tok/s] ` + thinking.slice(0, 200));
    } else {
      output.textContent = text || '(empty)';
      if (thinking) l('[thinking] ' + thinking.slice(0, 100));
      l(`[${tps} tok/s] ` + (text || '(empty)').slice(0, 200));
    }
    l(`Done in ${elapsed}s`);
    status.innerHTML = `<span class="green">*</span> done (${elapsed}s)`;
  } catch (e) {
    l('ERROR: ' + e.message);
    console.error(e);
    status.innerHTML = '<span class="red">*</span> error';
  }
  document.getElementById('btn-gen').disabled = false;
};
</script>
</body>
</html>