zero-tvm/wllama-bench.html at main · abgnydn/zero-tvm · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
<!DOCTYPE html>
<html lang="en">
<head>
  <meta charset="UTF-8">
  <meta name="viewport" content="width=device-width, initial-scale=1.0">
  <title>wllama Bench — llama.cpp WebGPU baseline vs Zero-TVM</title>
  <meta name="description" content="Third benchmark baseline: wllama (llama.cpp WASM + WebGPU backend) on GGUF Q4_K_M vs Zero-TVM on MLC q4f16_1. NOT a same-bytes comparison — the quantizations differ, so this measures runtime AND quantization.">
  <meta name="theme-color" content="#0a0a0c">
  <link rel="icon" href="/favicon.svg" type="image/svg+xml">
  <meta name="robots" content="noindex">
  <link rel="stylesheet" href="/style.css">
</head>
<body>
  <div class="container">
    <header>
      <a href="/" class="back-link">← zerotvm.com</a>
      <h1>wllama (llama.cpp WebGPU) Baseline</h1>
      <span id="badge" class="badge loading">Loading...</span>
      <span id="stats" class="stats"></span>
    </header>
    <p style="max-width:70ch;line-height:1.5">
      <strong>Not a same-bytes comparison.</strong> wllama reads GGUF
      (Q4_K_M / Q4); Zero-TVM and the WebLLM baseline read MLC
      <code>q4f16_1</code>. Same base model, <em>different quantization</em>.
      Every number on this page is therefore a <strong>runtime +
      quantization</strong> result, not the clean runtime-only A/B that
      <a href="/docs.html#vs-webllm">webllm-bench</a> is. A true same-bytes race
      against llama.cpp needs GGUF support in Zero-TVM.
    </p>
    <pre id="log" class="log"></pre>
  </div>
  <footer>wllama v3.5.1 (llama.cpp WASM + WebGPU backend) measured against Zero-TVM — different quantization, read the caveat above.</footer>
  <script type="module" src="/src/wllama-bench/main.ts"></script>
</body>
</html>